Skip to content

resource_tracker

resource_tracker #

Resource Tracker package for monitoring resource usage, detecting cloud environments, and more.

Modules:

Name Description
cloud_info

Detect cloud environment (provider, region, instance type) via VM metadata services.
dummy_workloads
helpers

General helpers.
keeper
nvidia

Helpers to monitor NVIDIA GPUs.
report
server_info

Detect server hardware (CPU count, memory amount, disk space, GPU count and VRAM amount) via procfs or psutil, and nvidia-smi.
tiny_bars

A tiny, partial, and opinionated implementation of the Handlebars template engine.
tiny_data_frame

A very inefficient data-frame implementation for manipulating resource usage data.
tracker

Track resource usage of a process and/or the system.
tracker_procfs

Helpers to track resource usage via procfs.
tracker_psutil

Helpers to track resource usage via psutil.

Classes:

Name Description
ProcessTracker

Track resource usage of a process and optionally its children.
ResourceTracker

Track resource usage of processes and the system in a non-blocking way.
SystemTracker

Track system-wide resource usage.

Functions:

Name Description
get_cloud_info

Detect cloud environment and return standardized information on provider, region, and instance type.
get_server_info

Collects important information about the Linux server.

get_cloud_info cached #

get_cloud_info()

Detect cloud environment and return standardized information on provider, region, and instance type.

Returns:

Type Description
dict

A dictionary containing standardized cloud information:

  • vendor: The cloud provider (aws, gcp, azure, hcloud, upcloud), or "unknown"
  • instance_type: The instance type/size/flavor, or "unknown"
  • region: The region/zone where the instance is running, or "unknown"
  • discovery_time: The time taken to discover the cloud environment, in seconds
Source code in resource_tracker/cloud_info.py 
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
@cache
def get_cloud_info() -> dict:
    """
    Detect cloud environment and return standardized information on provider, region, and instance type.

    Returns:
        A dictionary containing standardized cloud information:

            - `vendor`: The cloud provider (aws, gcp, azure, hcloud, upcloud), or "unknown"
            - `instance_type`: The instance type/size/flavor, or "unknown"
            - `region`: The region/zone where the instance is running, or "unknown"
            - `discovery_time`: The time taken to discover the cloud environment, in seconds
    """
    start_time = time()
    check_functions = [
        _check_aws,
        _check_gcp,
        _check_azure,
        _check_hetzner,
        _check_upcloud,
    ]

    # run checks in parallel, return early if any check succeeds
    with ThreadPoolExecutor(max_workers=len(check_functions)) as executor:
        futures = {executor.submit(check_fn): check_fn for check_fn in check_functions}
        pending = set(futures.keys())
        while pending:
            done, pending = wait(pending, return_when=FIRST_COMPLETED)
            for future in done:
                with suppress(Exception):
                    info = future.result()
                    if info:
                        # stop all remaining checks early
                        for f in pending:
                            f.cancel()
                        return info | {"discovery_time": time() - start_time}

    return {
        "vendor": "unknown",
        "instance_type": "unknown",
        "region": "unknown",
        "discovery_time": time() - start_time,
    }

get_server_info #

get_server_info()

Collects important information about the Linux server.

Returns:

Type Description
dict

A dictionary containing server information:

  • os: Operating system
  • vcpus: Number of virtual CPUs
  • memory_mb: Total memory in MB
  • gpu_count: Number of GPUs (0 if not available)
  • gpu_names: List of GPU names ([] if not available)
  • gpu_name: Most common GPU name ("" if not available)
  • gpu_memory_mb: Total VRAM in MB (0 if not available)
Source code in resource_tracker/server_info.py 
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
def get_server_info() -> dict:
    """
    Collects important information about the Linux server.

    Returns:
        A dictionary containing server information:

            - `os`: Operating system
            - `vcpus`: Number of virtual CPUs
            - `memory_mb`: Total memory in MB
            - `gpu_count`: Number of GPUs (`0` if not available)
            - `gpu_names`: List of GPU names (`[]` if not available)
            - `gpu_name`: Most common GPU name (`""` if not available)
            - `gpu_memory_mb`: Total VRAM in MB (`0` if not available)
    """
    gpu_info = get_gpu_info()
    info = {
        "os": system(),
        "vcpus": cpu_count(),
        "memory_mb": get_total_memory_mb(),
        "gpu_count": gpu_info["count"],
        "gpu_names": gpu_info["gpu_names"],
        "gpu_name": (
            Counter(gpu_info["gpu_names"]).most_common(1)[0][0]
            if gpu_info["gpu_names"]
            else ""
        ),
        "gpu_memory_mb": gpu_info["memory_mb"],
    }
    return info

ProcessTracker #

Track resource usage of a process and optionally its children.

This class monitors system resources like CPU times and usage, memory usage, GPU and VRAM utilization, I/O operations for a given process ID and optionally its child processes.

Data is collected every interval seconds and written to the stdout or output_file (if provided) as CSV. Currently, the following columns are tracked:

  • timestamp (float): The current timestamp.
  • pid (int): The monitored process ID.
  • children (int | None): The current number of child processes.
  • utime (int): The total user+nice mode CPU time in seconds.
  • stime (int): The total system mode CPU time in seconds.
  • cpu_usage (float): The current CPU usage between 0 and number of CPUs.
  • memory (int): The current memory usage in kB. Implementation depends on the operating system, and it is preferably PSS (Proportional Set Size) on Linux, USS (Unique Set Size) on macOS and Windows, and RSS (Resident Set Size) on Windows.
  • read_bytes (int): The total number of bytes read from disk.
  • write_bytes (int): The total number of bytes written to disk.
  • gpu_usage (float): The current GPU utilization between 0 and GPU count.
  • gpu_vram (float): The current GPU memory used in MiB.
  • gpu_utilized (int): The number of GPUs with utilization > 0.

Parameters:

Name Type Description Default
pid int

Process ID to track. Defaults to current process ID.

 getpid()
start_time float

Time when to start tracking. Defaults to current time.

 time()
interval float

Sampling interval in seconds. Defaults to 1.

 1
children bool

Whether to track child processes. Defaults to True.

 True
autostart bool

Whether to start tracking immediately. Defaults to True.

 True
output_file str

File to write the output to. Defaults to None, print to stdout.

 None

Methods:

Name Description
__call__

Dummy method to make this class callable.
diff_stats

Calculate stats since last call.
start_tracking

Start an infinite loop tracking resource usage of the process until it exits.
Source code in resource_tracker/tracker.py 
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
class ProcessTracker:
    """Track resource usage of a process and optionally its children.

    This class monitors system resources like CPU times and usage, memory usage,
    GPU and VRAM utilization, I/O operations for a given process ID and
    optionally its child processes.

    Data is collected every `interval` seconds and written to the stdout or
    `output_file` (if provided) as CSV. Currently, the following columns are
    tracked:

    - timestamp (float): The current timestamp.
    - pid (int): The monitored process ID.
    - children (int | None): The current number of child processes.
    - utime (int): The total user+nice mode CPU time in seconds.
    - stime (int): The total system mode CPU time in seconds.
    - cpu_usage (float): The current CPU usage between 0 and number of CPUs.
    - memory (int): The current memory usage in kB. Implementation depends on the
      operating system, and it is preferably PSS (Proportional Set Size) on Linux,
      USS (Unique Set Size) on macOS and Windows, and RSS (Resident Set Size) on
      Windows.
    - read_bytes (int): The total number of bytes read from disk.
    - write_bytes (int): The total number of bytes written to disk.
    - gpu_usage (float): The current GPU utilization between 0 and GPU count.
    - gpu_vram (float): The current GPU memory used in MiB.
    - gpu_utilized (int): The number of GPUs with utilization > 0.

    Args:
        pid (int, optional): Process ID to track. Defaults to current process ID.
        start_time: Time when to start tracking. Defaults to current time.
        interval (float, optional): Sampling interval in seconds. Defaults to 1.
        children (bool, optional): Whether to track child processes. Defaults to True.
        autostart (bool, optional): Whether to start tracking immediately. Defaults to True.
        output_file (str, optional): File to write the output to. Defaults to None, print to stdout.
    """

    def __init__(
        self,
        pid: int = getpid(),
        start_time: float = time(),
        interval: float = 1,
        children: bool = True,
        autostart: bool = True,
        output_file: str = None,
    ):
        self.get_process_stats, _ = get_tracker_implementation()

        self.pid = pid
        self.status = "running"
        self.interval = interval
        self.cycle = 0
        self.children = children
        self.start_time = start_time

        # initial data collection so that we can use that as a baseline when diffing after the first interval
        self.stats = self.get_process_stats(pid, children)

        if autostart:
            # wait for the start time to be reached
            if start_time > time():
                sleep(start_time - time())
            # we can now start. 1st interval used to collect baseline
            self.start_tracking(output_file)

    def __call__(self):
        """Dummy method to make this class callable."""
        pass

    def diff_stats(self):
        """Calculate stats since last call."""
        last_stats = self.stats
        self.stats = self.get_process_stats(self.pid, self.children)
        self.cycle += 1

        return {
            "timestamp": round(self.stats["timestamp"], 3),
            "pid": self.pid,
            "children": self.stats["children"],
            "utime": max(0, round(self.stats["utime"] - last_stats["utime"], 6)),
            "stime": max(0, round(self.stats["stime"] - last_stats["stime"], 6)),
            "cpu_usage": round(
                max(
                    0,
                    (
                        (self.stats["utime"] + self.stats["stime"])
                        - (last_stats["utime"] + last_stats["stime"])
                    )
                    / (self.stats["timestamp"] - last_stats["timestamp"]),
                ),
                4,
            ),
            "memory": self.stats["memory"],
            "read_bytes": max(0, self.stats["read_bytes"] - last_stats["read_bytes"]),
            "write_bytes": max(
                0, self.stats["write_bytes"] - last_stats["write_bytes"]
            ),
            "gpu_usage": self.stats["gpu_usage"],
            "gpu_vram": self.stats["gpu_vram"],
            "gpu_utilized": self.stats["gpu_utilized"],
        }

    def start_tracking(
        self, output_file: Optional[str] = None, print_header: bool = True
    ):
        """Start an infinite loop tracking resource usage of the process until it exits.

        A CSV line is written every `interval` seconds.

        Args:
            output_file: File to write the output to. Defaults to None, printing to stdout.
            print_header: Whether to print the header of the CSV. Defaults to True.
        """
        file_handle = open(output_file, "wb") if output_file else stdout.buffer
        try:
            while True:
                current_stats = self.diff_stats()
                if current_stats["memory"] == 0:
                    # the process has exited
                    self.status = "exited"
                    break
                # don't print values yet, we collect data for the 1st baseline
                if self.cycle == 1:
                    if print_header:
                        file_handle.write(
                            render_csv_row(
                                current_stats.keys(), quoting=QUOTE_NONNUMERIC
                            )
                        )
                else:
                    file_handle.write(
                        render_csv_row(current_stats.values(), quoting=QUOTE_NONNUMERIC)
                    )
                if output_file:
                    file_handle.flush()
                # sleep until the next interval
                sleep(max(0, self.start_time + self.interval * self.cycle - time()))
        finally:
            if output_file and not file_handle.closed:
                file_handle.close()

__call__ #

__call__()

Dummy method to make this class callable.

Source code in resource_tracker/tracker.py 
124
125
126
def __call__(self):
    """Dummy method to make this class callable."""
    pass

diff_stats #

diff_stats()

Calculate stats since last call.

Source code in resource_tracker/tracker.py 
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
def diff_stats(self):
    """Calculate stats since last call."""
    last_stats = self.stats
    self.stats = self.get_process_stats(self.pid, self.children)
    self.cycle += 1

    return {
        "timestamp": round(self.stats["timestamp"], 3),
        "pid": self.pid,
        "children": self.stats["children"],
        "utime": max(0, round(self.stats["utime"] - last_stats["utime"], 6)),
        "stime": max(0, round(self.stats["stime"] - last_stats["stime"], 6)),
        "cpu_usage": round(
            max(
                0,
                (
                    (self.stats["utime"] + self.stats["stime"])
                    - (last_stats["utime"] + last_stats["stime"])
                )
                / (self.stats["timestamp"] - last_stats["timestamp"]),
            ),
            4,
        ),
        "memory": self.stats["memory"],
        "read_bytes": max(0, self.stats["read_bytes"] - last_stats["read_bytes"]),
        "write_bytes": max(
            0, self.stats["write_bytes"] - last_stats["write_bytes"]
        ),
        "gpu_usage": self.stats["gpu_usage"],
        "gpu_vram": self.stats["gpu_vram"],
        "gpu_utilized": self.stats["gpu_utilized"],
    }

start_tracking #

start_tracking(output_file=None, print_header=True)

Start an infinite loop tracking resource usage of the process until it exits.

A CSV line is written every interval seconds.

Parameters:

Name Type Description Default
output_file Optional[str]

File to write the output to. Defaults to None, printing to stdout.

 None
print_header bool

Whether to print the header of the CSV. Defaults to True.

 True
Source code in resource_tracker/tracker.py 
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
def start_tracking(
    self, output_file: Optional[str] = None, print_header: bool = True
):
    """Start an infinite loop tracking resource usage of the process until it exits.

    A CSV line is written every `interval` seconds.

    Args:
        output_file: File to write the output to. Defaults to None, printing to stdout.
        print_header: Whether to print the header of the CSV. Defaults to True.
    """
    file_handle = open(output_file, "wb") if output_file else stdout.buffer
    try:
        while True:
            current_stats = self.diff_stats()
            if current_stats["memory"] == 0:
                # the process has exited
                self.status = "exited"
                break
            # don't print values yet, we collect data for the 1st baseline
            if self.cycle == 1:
                if print_header:
                    file_handle.write(
                        render_csv_row(
                            current_stats.keys(), quoting=QUOTE_NONNUMERIC
                        )
                    )
            else:
                file_handle.write(
                    render_csv_row(current_stats.values(), quoting=QUOTE_NONNUMERIC)
                )
            if output_file:
                file_handle.flush()
            # sleep until the next interval
            sleep(max(0, self.start_time + self.interval * self.cycle - time()))
    finally:
        if output_file and not file_handle.closed:
            file_handle.close()

ResourceTracker #

Track resource usage of processes and the system in a non-blocking way.

Start a resource_tracker.ProcessTracker and/or a resource_tracker.SystemTracker in the background as spawned or forked process(es), and make the collected data available easily in the main process via the process_metrics and system_metrics properties.

Parameters:

Name Type Description Default
pid int

Process ID to track. Defaults to current process ID.

 getpid()
children bool

Whether to track child processes. Defaults to True.

 True
interval float

Sampling interval in seconds. Defaults to 1.

 1
method Optional[str]

Multiprocessing method. Defaults to None, which tries to fork on Linux and macOS, and spawn on Windows.

 None
autostart bool

Whether to start tracking immediately. Defaults to True.

 True
track_processes bool

Whether to track resource usage at the process level. Defaults to True.

 True
track_system bool

Whether to track system-wide resource usage. Defaults to True.

 True
discover_server bool

Whether to discover the server specs in the background at startup. Defaults to True.

 True
discover_cloud bool

Whether to discover the cloud environment in the background at startup. Defaults to True.

 True

Example:

>>> from resource_tracker.dummy_workloads import cpu_single, cpu_multi
>>> tracker = ResourceTracker()
>>> cpu_single()
>>> tracker.recommend_resources()  # doctest: +SKIP
{'cpu': 1, 'memory': 128, 'gpu': 0, 'vram': 0}
>>> tracker = ResourceTracker()
>>> while tracker.n_samples == 0:
...     cpu_multi(duration=0.25, ncores=2)
>>> tracker.recommend_resources()  # multiprocessing is not enough efficient on Windows/macOS  # doctest: +SKIP
{'cpu': 2, 'memory': 128, 'gpu': 0, 'vram': 0}

Methods:

Name Description
start

Start the selected resource trackers in the background as subprocess(es).
cleanup

Cleanup temp files and background processes.
stop

Stop the previously started resource trackers' background processes.
snapshot

Collect the current state of the resource tracker.
from_snapshot

Create a ResourceTracker from a snapshot.
dumps

Serialize the resource tracker to a JSON string.
loads

Deserialize the resource tracker from a JSON string.
dump

Serialize the resource tracker to a gzipped JSON file.
load

Deserialize the resource tracker from a gzipped JSON file.
get_combined_metrics

Collected data both from the resource_tracker.ProcessTracker and resource_tracker.SystemTracker.
stats

Collect statistics from the resource tracker.
wait_for_samples

Wait for at least one sample to be collected.
recommend_resources

Recommend optimal resource allocation based on the measured resource tracker data.
recommend_server

Recommend the cheapest cloud server matching the recommended resources.

Attributes:

Name Type Description
n_samples int

Number of samples collected by the resource tracker.
server_info dict

High-level server info.
cloud_info dict

High-level cloud info.
process_metrics TinyDataFrame

Collected data from resource_tracker.ProcessTracker.
system_metrics TinyDataFrame

Collected data from resource_tracker.SystemTracker.
running bool

Check if the resource tracker is running.
Source code in resource_tracker/tracker.py 
 453
 454
 455
 456
 457
 458
 459
 460
 461
 462
 463
 464
 465
 466
 467
 468
 469
 470
 471
 472
 473
 474
 475
 476
 477
 478
 479
 480
 481
 482
 483
 484
 485
 486
 487
 488
 489
 490
 491
 492
 493
 494
 495
 496
 497
 498
 499
 500
 501
 502
 503
 504
 505
 506
 507
 508
 509
 510
 511
 512
 513
 514
 515
 516
 517
 518
 519
 520
 521
 522
 523
 524
 525
 526
 527
 528
 529
 530
 531
 532
 533
 534
 535
 536
 537
 538
 539
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 556
 557
 558
 559
 560
 561
 562
 563
 564
 565
 566
 567
 568
 569
 570
 571
 572
 573
 574
 575
 576
 577
 578
 579
 580
 581
 582
 583
 584
 585
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
class ResourceTracker:
    """Track resource usage of processes and the system in a non-blocking way.

    Start a [resource_tracker.ProcessTracker][] and/or a [resource_tracker.SystemTracker][] in the background as spawned
    or forked process(es), and make the collected data available easily in the
    main process via the `process_metrics` and `system_metrics` properties.

    Args:
        pid: Process ID to track. Defaults to current process ID.
        children: Whether to track child processes. Defaults to True.
        interval: Sampling interval in seconds. Defaults to 1.
        method: Multiprocessing method. Defaults to None, which tries to fork on
            Linux and macOS, and spawn on Windows.
        autostart: Whether to start tracking immediately. Defaults to True.
        track_processes: Whether to track resource usage at the process level.
            Defaults to True.
        track_system: Whether to track system-wide resource usage. Defaults to True.
        discover_server: Whether to discover the server specs in the background at
            startup. Defaults to True.
        discover_cloud: Whether to discover the cloud environment in the background
            at startup. Defaults to True.

    Example:

        >>> from resource_tracker.dummy_workloads import cpu_single, cpu_multi
        >>> tracker = ResourceTracker()
        >>> cpu_single()
        >>> tracker.recommend_resources()  # doctest: +SKIP
        {'cpu': 1, 'memory': 128, 'gpu': 0, 'vram': 0}
        >>> tracker = ResourceTracker()
        >>> while tracker.n_samples == 0:
        ...     cpu_multi(duration=0.25, ncores=2)
        >>> tracker.recommend_resources()  # multiprocessing is not enough efficient on Windows/macOS  # doctest: +SKIP
        {'cpu': 2, 'memory': 128, 'gpu': 0, 'vram': 0}
    """

    _server_info: Optional[dict] = None
    _cloud_info: Optional[dict] = None

    def __init__(
        self,
        pid: int = getpid(),
        children: bool = True,
        interval: float = 1,
        method: Optional[str] = None,
        autostart: bool = True,
        track_processes: bool = True,
        track_system: bool = True,
        discover_server: bool = True,
        discover_cloud: bool = True,
    ):
        self.pid = pid
        self.children = children
        self.interval = interval
        self.method = method
        self.autostart = autostart
        self.trackers = []
        if track_processes:
            self.trackers.append("process_tracker")
        if track_system:
            self.trackers.append("system_tracker")
        self.discover_server = discover_server
        self.discover_cloud = discover_cloud

        if platform != "linux" and not is_psutil_available():
            raise ImportError(
                "psutil is required for resource tracking on non-Linux platforms"
            )

        if method is None:
            # try to fork when possible due to leaked semaphores on older Python versions
            # see e.g. https://github.com/python/cpython/issues/90549
            if platform in ["linux", "darwin"]:
                self.mpc = get_context("fork")
            else:
                self.mpc = get_context("spawn")
        else:
            self.mpc = get_context(method)

        # error details from subprocesses
        self.error_queue = self.mpc.SimpleQueue()

        # create temporary CSV file(s) for the tracker(s), and record only the file path(s)
        # to be passed later to subprocess(es) avoiding pickling the file object(s)
        for tracker_name in self.trackers:
            temp_file = NamedTemporaryFile(delete=False)
            setattr(self, f"{tracker_name}_filepath", temp_file.name)
            temp_file.close()
        # make sure to cleanup the temp file(s)
        finalize(
            self,
            cleanup_files,
            [
                getattr(self, f"{tracker_name}_filepath")
                for tracker_name in self.trackers
            ],
        )

        if autostart:
            self.start()

    def start(self):
        """Start the selected resource trackers in the background as subprocess(es)."""
        if self.running:
            raise RuntimeError("Resource tracker already running, cannot start again.")
        if hasattr(self, "stop_time"):
            raise RuntimeError(
                "Resource tracker already stopped. Create a new instance instead of trying to restart it."
            )

        self.start_time = time()
        self.stop_time = None
        # round to the nearest interval in the future
        self.start_time = ceil(self.start_time / self.interval) * self.interval
        # leave at least 50 ms for trackers to start
        if self.start_time - time() < 0.05:
            self.start_time += self.interval

        if "process_tracker" in self.trackers:
            self.process_tracker_process = self.mpc.Process(
                target=_run_tracker,
                args=("process", self.error_queue),
                kwargs={
                    "pid": self.pid,
                    "start_time": self.start_time,
                    "interval": self.interval,
                    "children": self.children,
                    "output_file": self.process_tracker_filepath,
                },
                daemon=True,
            )
            self.process_tracker_process.start()

        if "system_tracker" in self.trackers:
            self.system_tracker_process = self.mpc.Process(
                target=_run_tracker,
                args=("system", self.error_queue),
                kwargs={
                    "start_time": self.start_time,
                    "interval": self.interval,
                    "output_file": self.system_tracker_filepath,
                },
                daemon=True,
            )
            self.system_tracker_process.start()

        def collect_server_info():
            """Collect server info to be run in a background thread."""
            try:
                self._server_info = get_server_info()
            except Exception as e:
                logger.warning(f"Error fetching server info: {e}")

        def collect_cloud_info():
            """Collect cloud info to be run in a background thread."""
            try:
                self._cloud_info = get_cloud_info()
            except Exception as e:
                logger.warning(f"Error fetching cloud info: {e}")

        if self.discover_server:
            server_thread = Thread(target=collect_server_info, daemon=True)
            server_thread.start()
        if self.discover_cloud:
            cloud_thread = Thread(target=collect_cloud_info, daemon=True)
            cloud_thread.start()

        # make sure to cleanup the started subprocess(es)
        finalize(
            self,
            cleanup_processes,
            [
                getattr(self, f"{tracker_name}_process")
                for tracker_name in self.trackers
            ],
        )

    def cleanup(self):
        """Cleanup temp files and background processes.

        Note that there is no need to call this method manually, as it is
        automatically handled by the garbage collector, but in some cases it
        might be useful to call it manually to avoid waiting for the garbage
        collector to run.
        """
        with suppress(Exception):
            self.stop()
        with suppress(Exception):
            cleanup_files(
                [
                    getattr(self, f"{tracker_name}_filepath")
                    for tracker_name in self.trackers
                ]
            )
        with suppress(Exception):
            cleanup_processes(
                [
                    getattr(self, f"{tracker_name}_process")
                    for tracker_name in self.trackers
                ]
            )

    def stop(self):
        """Stop the previously started resource trackers' background processes."""
        self.stop_time = time()
        # check for errors in the subprocesses
        if not self.error_queue.empty():
            error_data = self.error_queue.get()
            logger.warning(
                "Resource tracker subprocess failed!\n"
                f"Error type: {error_data['name']} (from module {error_data['module']})\n"
                f"Error message: {error_data['message']}\n"
                f"Original traceback:\n{error_data['traceback']}"
            )
        # terminate tracker processes
        for tracker_name in self.trackers:
            process_attr = f"{tracker_name}_process"
            if hasattr(self, process_attr):
                cleanup_processes([getattr(self, process_attr)])
        self.error_queue.close()
        logger.debug(
            "Resource tracker stopped after %s seconds, logging %d process-level and %d system-wide records",
            self.stop_time - self.start_time,
            len(self.process_metrics),
            len(self.system_metrics),
        )

    @property
    def n_samples(self) -> int:
        """Number of samples collected by the resource tracker."""
        return min(len(self.process_metrics), len(self.system_metrics))

    @property
    def server_info(self) -> dict:
        """High-level server info.

        Collected data from [resource_tracker.get_server_info][] plus a guess
        for the allocation type of the server: if it's dedicated to the tracked
        process(es) or shared with other processes. The guess is based on the
        [resource_tracker.column_maps.SERVER_ALLOCATION_CHECKS][] checks.
        """
        server_info = self._server_info
        if server_info:
            server_info["allocation"] = None
        if self.n_samples > 0:
            for check in SERVER_ALLOCATION_CHECKS:
                try:
                    system_val = mean(self.system_metrics[check["system_column"]])
                    task_val = mean(self.process_metrics[check["process_column"]])
                    if (system_val > task_val * check["percent"]) or (
                        system_val > task_val + check["absolute"]
                    ):
                        server_info["allocation"] = "shared"
                        break
                except Exception as e:
                    logger.warning(
                        f"Error calculating server allocation based on {check['system_column']} and {check['process_column']}: {e}"
                    )
                    server_info["allocation"] = "unknown"
                    break
            server_info["allocation"] = server_info.get("allocation", "dedicated")
        return server_info

    @property
    def cloud_info(self) -> dict:
        """High-level cloud info.

        Collected data from [resource_tracker.get_cloud_info][].
        """
        return self._cloud_info

    @property
    def process_metrics(self) -> TinyDataFrame:
        """Collected data from [resource_tracker.ProcessTracker][].

        Returns:
            A [resource_tracker.tiny_data_frame.TinyDataFrame][] object containing the collected data or an empty list if the [resource_tracker.ProcessTracker][] is not running.
        """
        try:
            return TinyDataFrame(
                csv_file_path=self.process_tracker_filepath,
                retries=2,
                retry_delay=min(0.05, self.interval / 10),
            )
        except Exception as e:
            logger.warning(f"Failed to read process metrics: {e}")
            return TinyDataFrame(data=[])

    @property
    def system_metrics(self) -> TinyDataFrame:
        """Collected data from [resource_tracker.SystemTracker][].

        Returns:
            A [resource_tracker.tiny_data_frame.TinyDataFrame][] object containing the collected data or an empty list if the [resource_tracker.SystemTracker][] is not running.
        """
        try:
            return TinyDataFrame(
                csv_file_path=self.system_tracker_filepath,
                retries=2,
                retry_delay=min(0.05, self.interval / 10),
            )
        except Exception as e:
            logger.warning(f"Failed to read system metrics: {e}")
            return TinyDataFrame(data=[])

    def snapshot(self) -> dict:
        """Collect the current state of the resource tracker.

        Returns:
            A dictionary containing the current state of the resource tracker.
        """
        return {
            "metadata": {
                "version": 1,
                "resource_tracker": {
                    "version": __version__,
                    "implementation": "psutil" if is_psutil_available() else "procfs",
                },
                "pid": self.pid,
                "children": self.children,
                "interval": self.interval,
                "method": self.method,
                "autostart": self.autostart,
                "track_processes": "process_tracker" in self.trackers,
                "track_system": "system_tracker" in self.trackers,
                "discover_server": self.discover_server,
                "discover_cloud": self.discover_cloud,
                "start_time": self.start_time,
                "stop_time": self.stop_time or time(),
                "duration": round(
                    (self.stop_time or time()) - self.start_time + self.interval, 2
                ),
            },
            "server_info": self.server_info,
            "cloud_info": self.cloud_info,
            "process_metrics": self.process_metrics.to_dict(),
            "system_metrics": self.system_metrics.to_dict(),
        }

    @classmethod
    def from_snapshot(cls, snapshot: dict):
        """Create a ResourceTracker from a snapshot.

        Args:
            snapshot: A dictionary containing the current state of the resource tracker, created by [resource_tracker.ResourceTracker.snapshot][].
        """
        tracker = cls(
            pid=snapshot["metadata"]["pid"],
            children=snapshot["metadata"]["children"],
            interval=snapshot["metadata"]["interval"],
            method=snapshot["metadata"]["method"],
            autostart=False,
            track_processes=snapshot["metadata"]["track_processes"],
            track_system=snapshot["metadata"]["track_system"],
            discover_server=snapshot["metadata"]["discover_server"],
            discover_cloud=snapshot["metadata"]["discover_cloud"],
        )
        tracker.start_time = snapshot["metadata"]["start_time"]
        tracker.stop_time = snapshot["metadata"]["stop_time"]
        tracker._server_info = snapshot["server_info"]
        tracker._cloud_info = snapshot["cloud_info"]
        TinyDataFrame(data=snapshot["process_metrics"]).to_csv(
            tracker.process_tracker_filepath
        )
        TinyDataFrame(data=snapshot["system_metrics"]).to_csv(
            tracker.system_tracker_filepath
        )
        return tracker

    def dumps(self) -> str:
        """Serialize the resource tracker to a JSON string.

        Returns:
            A JSON string containing the current state of the resource tracker.
        """
        return json_dumps(self.snapshot())

    @classmethod
    def loads(cls, s: str):
        """Deserialize the resource tracker from a JSON string.

        Args:
            s: The JSON string to deserialize the resource tracker from.
        """
        return cls.from_snapshot(json_loads(s))

    def dump(self, file: str):
        """Serialize the resource tracker to a gzipped JSON file.

        Args:
            file: The path to the file to write the serialized resource tracker to.
        """
        with gzip_open(file, "wb") as f:
            f.write(self.dumps().encode())

    @classmethod
    def load(cls, file: str):
        """Deserialize the resource tracker from a gzipped JSON file.

        Args:
            file: The path to the file to read the serialized resource tracker from.
        """
        with gzip_open(file, "rb") as f:
            return cls.loads(f.read().decode())

    def get_combined_metrics(
        self,
        bytes: bool = False,
        human_names: bool = False,
        system_prefix: Optional[str] = None,
        process_prefix: Optional[str] = None,
    ) -> TinyDataFrame:
        """Collected data both from the [resource_tracker.ProcessTracker][] and [resource_tracker.SystemTracker][].

        This is effectively binding the two dataframes together by timestamp,
        and adding a prefix to the column names to distinguish between the system and process metrics.

        Args:
            bytes: Whether to convert all metrics (e.g. memory, VRAM, disk usage) to bytes. Defaults to False, reporting as documented at [resource_tracker.ProcessTracker][] and [resource_tracker.SystemTracker][] (kB, MiB, or GiB).
            human_names: Whether to rename the columns to use human-friendly names. Defaults to False, reporting as documented at [resource_tracker.ProcessTracker][] and [resource_tracker.SystemTracker][] with prefixes.
            system_prefix: Prefix to add to the system-level column names. Defaults to "system_" or "System " based on the value of `human_names`.
            process_prefix: Prefix to add to the process-level column names. Defaults to "process_" or "Process " based on the value of `human_names`.

        Returns:
            A [resource_tracker.tiny_data_frame.TinyDataFrame][] object containing the combined data or an empty list if tracker(s) not running.
        """
        try:
            process_metrics = self.process_metrics
            system_metrics = self.system_metrics

            # ensure both have the same length
            if len(process_metrics) > len(system_metrics):
                process_metrics = process_metrics[: len(system_metrics)]
            elif len(system_metrics) > len(process_metrics):
                system_metrics = system_metrics[: len(process_metrics)]

            # nothing to report on
            if len(process_metrics) == 0:
                return TinyDataFrame(data=[])

            if bytes:
                for col, factor in BYTE_MAPPING.items():
                    for metrics in (system_metrics, process_metrics):
                        if col in metrics.columns:
                            metrics[col] = [v * factor for v in metrics[col]]

            if system_prefix is None:
                system_prefix = "system_" if not human_names else "System "
            if process_prefix is None:
                process_prefix = "process_" if not human_names else "Process "

            # cbind the two dataframes with column name prefixes and optional human-friendly names
            combined = system_metrics.rename(
                columns={
                    n: (
                        (system_prefix if n != "timestamp" else "")
                        + (n if not human_names else HUMAN_NAMES_MAPPING.get(n, n))
                    )
                    for n in system_metrics.columns
                }
            )
            for col in process_metrics.columns[1:]:
                combined[
                    process_prefix
                    + (col if not human_names else HUMAN_NAMES_MAPPING.get(col, col))
                ] = process_metrics[col]

            return combined
        except Exception as e:
            with suppress(Exception):
                logger.warning(
                    f"Kept {len(process_metrics) if 'process_metrics' in locals() else 'unknown'} records of process metrics out of {len(self.process_metrics)} collected records ({self.process_metrics.columns}), "
                    f"and {len(system_metrics) if 'system_metrics' in locals() else 'unknown'} records of system metrics out of {len(self.system_metrics)} collected records ({self.system_metrics.columns}), "
                    f"but creating the combined metrics dataframe failed with error: {e}"
                )
                logger.warning(f"Process metrics: {self.process_metrics.to_dict()}")
                logger.warning(f"System metrics: {self.system_metrics.to_dict()}")
            raise

    def stats(
        self,
        specs: List[StatSpec] = [
            StatSpec(column="process_cpu_usage", agg=mean, round=2),
            StatSpec(column="process_cpu_usage", agg=max, round=2),
            StatSpec(column="process_memory", agg=mean, round=2),
            StatSpec(column="process_memory", agg=max, round=2),
            StatSpec(column="process_gpu_usage", agg=mean, round=2),
            StatSpec(column="process_gpu_usage", agg=max, round=2),
            StatSpec(column="process_gpu_vram", agg=mean, round=2),
            StatSpec(column="process_gpu_vram", agg=max, round=2),
            StatSpec(column="process_gpu_utilized", agg=mean, round=2),
            StatSpec(column="process_gpu_utilized", agg=max, round=2),
            StatSpec(column="system_disk_space_used_gb", agg=max, round=2),
            StatSpec(column="system_net_recv_bytes", agg=sum),
            StatSpec(column="system_net_sent_bytes", agg=sum),
            StatSpec(
                column="timestamp", agg=lambda x: max(x) - min(x), agg_name="duration"
            ),
        ],
    ) -> dict:
        """Collect statistics from the resource tracker.

        Args:
            specs: A list of [resource_tracker.tiny_data_frame.StatSpec][] objects specifying the statistics to collect.

        Returns:
            A dictionary containing the collected statistics.
        """
        if self.n_samples > 0:
            stats = self.get_combined_metrics().stats(specs)
            stats["timestamp"]["duration"] += self.interval
            return stats
        else:
            raise RuntimeError("No metrics collected (yet)")

    @property
    def running(self) -> bool:
        """Check if the resource tracker is running.

        Returns:
            True if the resource tracker is running, False if already stopped.
        """
        return hasattr(self, "stop_time") and self.stop_time is None

    def wait_for_samples(self, n: int = 1, timeout: float = 5):
        """Wait for at least one sample to be collected.

        Args:
            n: The minimum number of samples to collect. Defaults to 1.
            timeout: The maximum time to wait for a sample. Defaults to 5 seconds.
        """
        if self.running:
            while self.n_samples < n:
                sleep(self.interval / 10)
                if time() - self.start_time > timeout:
                    raise RuntimeError(
                        f"Timed out waiting for resource tracker to collect {n} samples"
                    )
        else:
            if self.n_samples < n:
                raise RuntimeError(
                    f"Resource tracker has been already stopped with {self.n_samples} sample(s), "
                    f"cannot wait to collect the requested {n} sample(s)."
                )

    def recommend_resources(self, historical_stats: List[dict] = []) -> dict:
        """Recommend optimal resource allocation based on the measured resource tracker data.

        The recommended resources are based on the following rules:

        - target average CPU usage of the process(es)
        - target maximum memory usage of the process(es) with a 20% buffer
        - target maximum number of GPUs used by the process(es)
        - target maximum VRAM usage of the process(es) with a 20% buffer

        Args:
            historical_stats: Optional list of historical statistics (as returned by [resource_tracker.ResourceTracker.stats][])
                              to consider when making recommendations. These will be combined with the current stats.

        Returns:
            A dictionary containing the recommended resources (cpu, memory, gpu, vram).
        """
        self.wait_for_samples(n=1, timeout=self.interval * 5)

        current_stats = self.stats()
        if historical_stats:
            stats = aggregate_stats([current_stats] + historical_stats)
        else:
            stats = current_stats

        rec = {}
        # target average CPU usage
        rec["cpu"] = max(1, round(stats["process_cpu_usage"]["mean"]))
        # target maximum memory usage (kB->MB) with a 20% buffer
        rec["memory"] = round_memory(mb=stats["process_memory"]["max"] * 1.2 / 1024)
        # target maximum GPU number of GPUs used
        rec["gpu"] = (
            max(1, round(stats["process_gpu_usage"]["max"]))
            if stats["process_gpu_usage"]["mean"] > 0
            else 0
        )
        # target maximum VRAM usage (MiB) with a 20% buffer
        rec["vram"] = (
            round_memory(mb=stats["process_gpu_vram"]["max"] * 1.2)
            if stats["process_gpu_vram"]["max"] > 0
            else 0
        )
        return rec

    def recommend_server(self, **kwargs) -> dict:
        """Recommend the cheapest cloud server matching the recommended resources.

        Args:
            **kwargs: Additional filtering arguments (e.g. vendor_id or compliance_framework_id) to pass to the Spare Cores Keeper API.

        Returns:
            A dictionary containing the recommended cloud server. Response format is described at <https://keeper.sparecores.net/redoc#tag/Query-Resources/operation/search_servers_servers_get>.
        """
        historical_stats = kwargs.pop("historical_stats", [])
        rec = self.recommend_resources(historical_stats=historical_stats)
        return get_recommended_cloud_servers(**rec, **kwargs, n=1)[0]

    def report(
        self,
        integration: Literal["standalone", "Metaflow", "R"] = "standalone",
        historical_stats: List[dict] = [],
        status_failed: bool = False,
        integration_version: Optional[str] = None,
    ) -> Report:
        self.wait_for_samples(n=1, timeout=self.interval * 5)
        duration = (self.stop_time or time()) - self.start_time + self.interval

        current_stats = self.stats()
        if historical_stats:
            combined_stats = aggregate_stats([current_stats] + historical_stats)
        else:
            combined_stats = current_stats

        ctx = {
            "files": _read_report_template_files(),
            "server_info": self.server_info,
            "cloud_info": self.cloud_info,
            "process_metrics": self.process_metrics,
            "system_metrics": self.system_metrics,
            "stats": current_stats,
            "historical_stats": historical_stats,
            "combined_stats": combined_stats,
            "recommended_resources": self.recommend_resources(
                historical_stats=historical_stats
            ),
            "recommended_server": self.recommend_server(
                historical_stats=historical_stats
            ),
            "resource_tracker": {
                "version": __version__,
                "implementation": "psutil" if is_psutil_available() else "procfs",
                "integration": integration,
                "integration_is": {
                    "standalone": integration == "standalone",
                    "not_standalone": integration != "standalone",
                    "metaflow": integration == "Metaflow",
                    "r": integration == "R",
                },
                "duration": duration,
                "start_time": self.start_time,
                "stop_time": self.stop_time,
                "stopped": self.stop_time is not None,
                "report_time": time(),
            },
            "status_failed": status_failed,
            "csv": {},
        }

        # comma-separated values
        joined = self.get_combined_metrics(bytes=True, human_names=True)
        for name, columns in REPORT_CSV_MAPPING.items():
            csv_data = joined[columns]
            # convert to JS milliseconds
            csv_data["Timestamp"] = [t * 1000 for t in csv_data["Timestamp"]]
            ctx["csv"][name] = csv_data.to_csv(quote_strings=False)

        # lookup instance prices
        rec_server_cost = ctx["recommended_server"]["min_price_ondemand"]
        rec_run_cost = rec_server_cost / 60 / 60 * duration
        ctx["recommended_server"]["best_ondemand_price_duration"] = rec_run_cost
        if ctx["cloud_info"] and ctx["cloud_info"]["instance_type"] != "unknown":
            current_server_cost = get_instance_price(
                ctx["cloud_info"]["vendor"],
                ctx["cloud_info"]["region"],
                ctx["cloud_info"]["instance_type"],
            )
            if current_server_cost:
                current_run_cost = round(current_server_cost / 60 / 60 * duration, 6)
                ctx["cloud_info"]["run_costs"] = current_run_cost
                ctx["recommended_server"]["cost_savings"] = {
                    "percent": round(
                        (current_run_cost - rec_run_cost) / current_run_cost * 100, 2
                    ),
                    "amount": round(current_run_cost - rec_run_cost, 6),
                }

        html_template_path = path.join(
            path.dirname(__file__), "report_template", "report.html"
        )
        with open(html_template_path) as f:
            html_template = f.read()
        html = render_template(html_template, ctx)
        return Report(html)

start #

start()

Start the selected resource trackers in the background as subprocess(es).

Source code in resource_tracker/tracker.py 
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
def start(self):
    """Start the selected resource trackers in the background as subprocess(es)."""
    if self.running:
        raise RuntimeError("Resource tracker already running, cannot start again.")
    if hasattr(self, "stop_time"):
        raise RuntimeError(
            "Resource tracker already stopped. Create a new instance instead of trying to restart it."
        )

    self.start_time = time()
    self.stop_time = None
    # round to the nearest interval in the future
    self.start_time = ceil(self.start_time / self.interval) * self.interval
    # leave at least 50 ms for trackers to start
    if self.start_time - time() < 0.05:
        self.start_time += self.interval

    if "process_tracker" in self.trackers:
        self.process_tracker_process = self.mpc.Process(
            target=_run_tracker,
            args=("process", self.error_queue),
            kwargs={
                "pid": self.pid,
                "start_time": self.start_time,
                "interval": self.interval,
                "children": self.children,
                "output_file": self.process_tracker_filepath,
            },
            daemon=True,
        )
        self.process_tracker_process.start()

    if "system_tracker" in self.trackers:
        self.system_tracker_process = self.mpc.Process(
            target=_run_tracker,
            args=("system", self.error_queue),
            kwargs={
                "start_time": self.start_time,
                "interval": self.interval,
                "output_file": self.system_tracker_filepath,
            },
            daemon=True,
        )
        self.system_tracker_process.start()

    def collect_server_info():
        """Collect server info to be run in a background thread."""
        try:
            self._server_info = get_server_info()
        except Exception as e:
            logger.warning(f"Error fetching server info: {e}")

    def collect_cloud_info():
        """Collect cloud info to be run in a background thread."""
        try:
            self._cloud_info = get_cloud_info()
        except Exception as e:
            logger.warning(f"Error fetching cloud info: {e}")

    if self.discover_server:
        server_thread = Thread(target=collect_server_info, daemon=True)
        server_thread.start()
    if self.discover_cloud:
        cloud_thread = Thread(target=collect_cloud_info, daemon=True)
        cloud_thread.start()

    # make sure to cleanup the started subprocess(es)
    finalize(
        self,
        cleanup_processes,
        [
            getattr(self, f"{tracker_name}_process")
            for tracker_name in self.trackers
        ],
    )

cleanup #

cleanup()

Cleanup temp files and background processes.

Note that there is no need to call this method manually, as it is automatically handled by the garbage collector, but in some cases it might be useful to call it manually to avoid waiting for the garbage collector to run.

Source code in resource_tracker/tracker.py 
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
def cleanup(self):
    """Cleanup temp files and background processes.

    Note that there is no need to call this method manually, as it is
    automatically handled by the garbage collector, but in some cases it
    might be useful to call it manually to avoid waiting for the garbage
    collector to run.
    """
    with suppress(Exception):
        self.stop()
    with suppress(Exception):
        cleanup_files(
            [
                getattr(self, f"{tracker_name}_filepath")
                for tracker_name in self.trackers
            ]
        )
    with suppress(Exception):
        cleanup_processes(
            [
                getattr(self, f"{tracker_name}_process")
                for tracker_name in self.trackers
            ]
        )

stop #

stop()

Stop the previously started resource trackers' background processes.

Source code in resource_tracker/tracker.py 
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
def stop(self):
    """Stop the previously started resource trackers' background processes."""
    self.stop_time = time()
    # check for errors in the subprocesses
    if not self.error_queue.empty():
        error_data = self.error_queue.get()
        logger.warning(
            "Resource tracker subprocess failed!\n"
            f"Error type: {error_data['name']} (from module {error_data['module']})\n"
            f"Error message: {error_data['message']}\n"
            f"Original traceback:\n{error_data['traceback']}"
        )
    # terminate tracker processes
    for tracker_name in self.trackers:
        process_attr = f"{tracker_name}_process"
        if hasattr(self, process_attr):
            cleanup_processes([getattr(self, process_attr)])
    self.error_queue.close()
    logger.debug(
        "Resource tracker stopped after %s seconds, logging %d process-level and %d system-wide records",
        self.stop_time - self.start_time,
        len(self.process_metrics),
        len(self.system_metrics),
    )

n_samples property #

n_samples

Number of samples collected by the resource tracker.

server_info property #

server_info

High-level server info.

Collected data from resource_tracker.get_server_info plus a guess for the allocation type of the server: if it's dedicated to the tracked process(es) or shared with other processes. The guess is based on the [resource_tracker.column_maps.SERVER_ALLOCATION_CHECKS][] checks.

cloud_info property #

cloud_info

High-level cloud info.

Collected data from resource_tracker.get_cloud_info.

process_metrics property #

process_metrics

Collected data from resource_tracker.ProcessTracker.

Returns:

Type Description
TinyDataFrame

A resource_tracker.tiny_data_frame.TinyDataFrame object containing the collected data or an empty list if the resource_tracker.ProcessTracker is not running.

system_metrics property #

system_metrics

Collected data from resource_tracker.SystemTracker.

Returns:

Type Description
TinyDataFrame

A resource_tracker.tiny_data_frame.TinyDataFrame object containing the collected data or an empty list if the resource_tracker.SystemTracker is not running.

snapshot #

snapshot()

Collect the current state of the resource tracker.

Returns:

Type Description
dict

A dictionary containing the current state of the resource tracker.
Source code in resource_tracker/tracker.py 
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
def snapshot(self) -> dict:
    """Collect the current state of the resource tracker.

    Returns:
        A dictionary containing the current state of the resource tracker.
    """
    return {
        "metadata": {
            "version": 1,
            "resource_tracker": {
                "version": __version__,
                "implementation": "psutil" if is_psutil_available() else "procfs",
            },
            "pid": self.pid,
            "children": self.children,
            "interval": self.interval,
            "method": self.method,
            "autostart": self.autostart,
            "track_processes": "process_tracker" in self.trackers,
            "track_system": "system_tracker" in self.trackers,
            "discover_server": self.discover_server,
            "discover_cloud": self.discover_cloud,
            "start_time": self.start_time,
            "stop_time": self.stop_time or time(),
            "duration": round(
                (self.stop_time or time()) - self.start_time + self.interval, 2
            ),
        },
        "server_info": self.server_info,
        "cloud_info": self.cloud_info,
        "process_metrics": self.process_metrics.to_dict(),
        "system_metrics": self.system_metrics.to_dict(),
    }

from_snapshot classmethod #

from_snapshot(snapshot)

Create a ResourceTracker from a snapshot.

Parameters:

Name Type Description Default
snapshot dict

A dictionary containing the current state of the resource tracker, created by resource_tracker.ResourceTracker.snapshot.

 required
Source code in resource_tracker/tracker.py 
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
@classmethod
def from_snapshot(cls, snapshot: dict):
    """Create a ResourceTracker from a snapshot.

    Args:
        snapshot: A dictionary containing the current state of the resource tracker, created by [resource_tracker.ResourceTracker.snapshot][].
    """
    tracker = cls(
        pid=snapshot["metadata"]["pid"],
        children=snapshot["metadata"]["children"],
        interval=snapshot["metadata"]["interval"],
        method=snapshot["metadata"]["method"],
        autostart=False,
        track_processes=snapshot["metadata"]["track_processes"],
        track_system=snapshot["metadata"]["track_system"],
        discover_server=snapshot["metadata"]["discover_server"],
        discover_cloud=snapshot["metadata"]["discover_cloud"],
    )
    tracker.start_time = snapshot["metadata"]["start_time"]
    tracker.stop_time = snapshot["metadata"]["stop_time"]
    tracker._server_info = snapshot["server_info"]
    tracker._cloud_info = snapshot["cloud_info"]
    TinyDataFrame(data=snapshot["process_metrics"]).to_csv(
        tracker.process_tracker_filepath
    )
    TinyDataFrame(data=snapshot["system_metrics"]).to_csv(
        tracker.system_tracker_filepath
    )
    return tracker

dumps #

dumps()

Serialize the resource tracker to a JSON string.

Returns:

Type Description
str

A JSON string containing the current state of the resource tracker.
Source code in resource_tracker/tracker.py 
822
823
824
825
826
827
828
def dumps(self) -> str:
    """Serialize the resource tracker to a JSON string.

    Returns:
        A JSON string containing the current state of the resource tracker.
    """
    return json_dumps(self.snapshot())

loads classmethod #

loads(s)

Deserialize the resource tracker from a JSON string.

Parameters:

Name Type Description Default
s str

The JSON string to deserialize the resource tracker from.

 required
Source code in resource_tracker/tracker.py 
830
831
832
833
834
835
836
837
@classmethod
def loads(cls, s: str):
    """Deserialize the resource tracker from a JSON string.

    Args:
        s: The JSON string to deserialize the resource tracker from.
    """
    return cls.from_snapshot(json_loads(s))

dump #

dump(file)

Serialize the resource tracker to a gzipped JSON file.

Parameters:

Name Type Description Default
file str

The path to the file to write the serialized resource tracker to.

 required
Source code in resource_tracker/tracker.py 
839
840
841
842
843
844
845
846
def dump(self, file: str):
    """Serialize the resource tracker to a gzipped JSON file.

    Args:
        file: The path to the file to write the serialized resource tracker to.
    """
    with gzip_open(file, "wb") as f:
        f.write(self.dumps().encode())

load classmethod #

load(file)

Deserialize the resource tracker from a gzipped JSON file.

Parameters:

Name Type Description Default
file str

The path to the file to read the serialized resource tracker from.

 required
Source code in resource_tracker/tracker.py 
848
849
850
851
852
853
854
855
856
@classmethod
def load(cls, file: str):
    """Deserialize the resource tracker from a gzipped JSON file.

    Args:
        file: The path to the file to read the serialized resource tracker from.
    """
    with gzip_open(file, "rb") as f:
        return cls.loads(f.read().decode())

get_combined_metrics #

get_combined_metrics(bytes=False, human_names=False, system_prefix=None, process_prefix=None)

Collected data both from the resource_tracker.ProcessTracker and resource_tracker.SystemTracker.

This is effectively binding the two dataframes together by timestamp, and adding a prefix to the column names to distinguish between the system and process metrics.

Parameters:

Name Type Description Default
bytes bool

Whether to convert all metrics (e.g. memory, VRAM, disk usage) to bytes. Defaults to False, reporting as documented at resource_tracker.ProcessTracker and resource_tracker.SystemTracker (kB, MiB, or GiB).

 False
human_names bool

Whether to rename the columns to use human-friendly names. Defaults to False, reporting as documented at resource_tracker.ProcessTracker and resource_tracker.SystemTracker with prefixes.

 False
system_prefix Optional[str]

Prefix to add to the system-level column names. Defaults to "system_" or "System " based on the value of human_names.

 None
process_prefix Optional[str]

Prefix to add to the process-level column names. Defaults to "process_" or "Process " based on the value of human_names.

 None

Returns:

Type Description
TinyDataFrame

A resource_tracker.tiny_data_frame.TinyDataFrame object containing the combined data or an empty list if tracker(s) not running.
Source code in resource_tracker/tracker.py 
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
def get_combined_metrics(
    self,
    bytes: bool = False,
    human_names: bool = False,
    system_prefix: Optional[str] = None,
    process_prefix: Optional[str] = None,
) -> TinyDataFrame:
    """Collected data both from the [resource_tracker.ProcessTracker][] and [resource_tracker.SystemTracker][].

    This is effectively binding the two dataframes together by timestamp,
    and adding a prefix to the column names to distinguish between the system and process metrics.

    Args:
        bytes: Whether to convert all metrics (e.g. memory, VRAM, disk usage) to bytes. Defaults to False, reporting as documented at [resource_tracker.ProcessTracker][] and [resource_tracker.SystemTracker][] (kB, MiB, or GiB).
        human_names: Whether to rename the columns to use human-friendly names. Defaults to False, reporting as documented at [resource_tracker.ProcessTracker][] and [resource_tracker.SystemTracker][] with prefixes.
        system_prefix: Prefix to add to the system-level column names. Defaults to "system_" or "System " based on the value of `human_names`.
        process_prefix: Prefix to add to the process-level column names. Defaults to "process_" or "Process " based on the value of `human_names`.

    Returns:
        A [resource_tracker.tiny_data_frame.TinyDataFrame][] object containing the combined data or an empty list if tracker(s) not running.
    """
    try:
        process_metrics = self.process_metrics
        system_metrics = self.system_metrics

        # ensure both have the same length
        if len(process_metrics) > len(system_metrics):
            process_metrics = process_metrics[: len(system_metrics)]
        elif len(system_metrics) > len(process_metrics):
            system_metrics = system_metrics[: len(process_metrics)]

        # nothing to report on
        if len(process_metrics) == 0:
            return TinyDataFrame(data=[])

        if bytes:
            for col, factor in BYTE_MAPPING.items():
                for metrics in (system_metrics, process_metrics):
                    if col in metrics.columns:
                        metrics[col] = [v * factor for v in metrics[col]]

        if system_prefix is None:
            system_prefix = "system_" if not human_names else "System "
        if process_prefix is None:
            process_prefix = "process_" if not human_names else "Process "

        # cbind the two dataframes with column name prefixes and optional human-friendly names
        combined = system_metrics.rename(
            columns={
                n: (
                    (system_prefix if n != "timestamp" else "")
                    + (n if not human_names else HUMAN_NAMES_MAPPING.get(n, n))
                )
                for n in system_metrics.columns
            }
        )
        for col in process_metrics.columns[1:]:
            combined[
                process_prefix
                + (col if not human_names else HUMAN_NAMES_MAPPING.get(col, col))
            ] = process_metrics[col]

        return combined
    except Exception as e:
        with suppress(Exception):
            logger.warning(
                f"Kept {len(process_metrics) if 'process_metrics' in locals() else 'unknown'} records of process metrics out of {len(self.process_metrics)} collected records ({self.process_metrics.columns}), "
                f"and {len(system_metrics) if 'system_metrics' in locals() else 'unknown'} records of system metrics out of {len(self.system_metrics)} collected records ({self.system_metrics.columns}), "
                f"but creating the combined metrics dataframe failed with error: {e}"
            )
            logger.warning(f"Process metrics: {self.process_metrics.to_dict()}")
            logger.warning(f"System metrics: {self.system_metrics.to_dict()}")
        raise

stats #

stats(specs=[StatSpec(column='process_cpu_usage', agg=mean, round=2), StatSpec(column='process_cpu_usage', agg=max, round=2), StatSpec(column='process_memory', agg=mean, round=2), StatSpec(column='process_memory', agg=max, round=2), StatSpec(column='process_gpu_usage', agg=mean, round=2), StatSpec(column='process_gpu_usage', agg=max, round=2), StatSpec(column='process_gpu_vram', agg=mean, round=2), StatSpec(column='process_gpu_vram', agg=max, round=2), StatSpec(column='process_gpu_utilized', agg=mean, round=2), StatSpec(column='process_gpu_utilized', agg=max, round=2), StatSpec(column='system_disk_space_used_gb', agg=max, round=2), StatSpec(column='system_net_recv_bytes', agg=sum), StatSpec(column='system_net_sent_bytes', agg=sum), StatSpec(column='timestamp', agg=lambda x: max(x) - min(x), agg_name='duration')])

Collect statistics from the resource tracker.

Parameters:

Name Type Description Default
specs List[StatSpec]

A list of resource_tracker.tiny_data_frame.StatSpec objects specifying the statistics to collect.

 [StatSpec(column='process_cpu_usage', agg=mean, round=2), StatSpec(column='process_cpu_usage', agg=max, round=2), StatSpec(column='process_memory', agg=mean, round=2), StatSpec(column='process_memory', agg=max, round=2), StatSpec(column='process_gpu_usage', agg=mean, round=2), StatSpec(column='process_gpu_usage', agg=max, round=2), StatSpec(column='process_gpu_vram', agg=mean, round=2), StatSpec(column='process_gpu_vram', agg=max, round=2), StatSpec(column='process_gpu_utilized', agg=mean, round=2), StatSpec(column='process_gpu_utilized', agg=max, round=2), StatSpec(column='system_disk_space_used_gb', agg=max, round=2), StatSpec(column='system_net_recv_bytes', agg=sum), StatSpec(column='system_net_sent_bytes', agg=sum), StatSpec(column='timestamp', agg=lambda x: max(x) - min(x), agg_name='duration')]

Returns:

Type Description
dict

A dictionary containing the collected statistics.
Source code in resource_tracker/tracker.py 
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
def stats(
    self,
    specs: List[StatSpec] = [
        StatSpec(column="process_cpu_usage", agg=mean, round=2),
        StatSpec(column="process_cpu_usage", agg=max, round=2),
        StatSpec(column="process_memory", agg=mean, round=2),
        StatSpec(column="process_memory", agg=max, round=2),
        StatSpec(column="process_gpu_usage", agg=mean, round=2),
        StatSpec(column="process_gpu_usage", agg=max, round=2),
        StatSpec(column="process_gpu_vram", agg=mean, round=2),
        StatSpec(column="process_gpu_vram", agg=max, round=2),
        StatSpec(column="process_gpu_utilized", agg=mean, round=2),
        StatSpec(column="process_gpu_utilized", agg=max, round=2),
        StatSpec(column="system_disk_space_used_gb", agg=max, round=2),
        StatSpec(column="system_net_recv_bytes", agg=sum),
        StatSpec(column="system_net_sent_bytes", agg=sum),
        StatSpec(
            column="timestamp", agg=lambda x: max(x) - min(x), agg_name="duration"
        ),
    ],
) -> dict:
    """Collect statistics from the resource tracker.

    Args:
        specs: A list of [resource_tracker.tiny_data_frame.StatSpec][] objects specifying the statistics to collect.

    Returns:
        A dictionary containing the collected statistics.
    """
    if self.n_samples > 0:
        stats = self.get_combined_metrics().stats(specs)
        stats["timestamp"]["duration"] += self.interval
        return stats
    else:
        raise RuntimeError("No metrics collected (yet)")

running property #

running

Check if the resource tracker is running.

Returns:

Type Description
bool

True if the resource tracker is running, False if already stopped.

wait_for_samples #

wait_for_samples(n=1, timeout=5)

Wait for at least one sample to be collected.

Parameters:

Name Type Description Default
n int

The minimum number of samples to collect. Defaults to 1.

 1
timeout float

The maximum time to wait for a sample. Defaults to 5 seconds.

 5
Source code in resource_tracker/tracker.py 
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
def wait_for_samples(self, n: int = 1, timeout: float = 5):
    """Wait for at least one sample to be collected.

    Args:
        n: The minimum number of samples to collect. Defaults to 1.
        timeout: The maximum time to wait for a sample. Defaults to 5 seconds.
    """
    if self.running:
        while self.n_samples < n:
            sleep(self.interval / 10)
            if time() - self.start_time > timeout:
                raise RuntimeError(
                    f"Timed out waiting for resource tracker to collect {n} samples"
                )
    else:
        if self.n_samples < n:
            raise RuntimeError(
                f"Resource tracker has been already stopped with {self.n_samples} sample(s), "
                f"cannot wait to collect the requested {n} sample(s)."
            )

recommend_resources #

recommend_resources(historical_stats=[])

Recommend optimal resource allocation based on the measured resource tracker data.

The recommended resources are based on the following rules:

  • target average CPU usage of the process(es)
  • target maximum memory usage of the process(es) with a 20% buffer
  • target maximum number of GPUs used by the process(es)
  • target maximum VRAM usage of the process(es) with a 20% buffer

Parameters:

Name Type Description Default
historical_stats List[dict]

Optional list of historical statistics (as returned by resource_tracker.ResourceTracker.stats) to consider when making recommendations. These will be combined with the current stats.

 []

Returns:

Type Description
dict

A dictionary containing the recommended resources (cpu, memory, gpu, vram).
Source code in resource_tracker/tracker.py 
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
def recommend_resources(self, historical_stats: List[dict] = []) -> dict:
    """Recommend optimal resource allocation based on the measured resource tracker data.

    The recommended resources are based on the following rules:

    - target average CPU usage of the process(es)
    - target maximum memory usage of the process(es) with a 20% buffer
    - target maximum number of GPUs used by the process(es)
    - target maximum VRAM usage of the process(es) with a 20% buffer

    Args:
        historical_stats: Optional list of historical statistics (as returned by [resource_tracker.ResourceTracker.stats][])
                          to consider when making recommendations. These will be combined with the current stats.

    Returns:
        A dictionary containing the recommended resources (cpu, memory, gpu, vram).
    """
    self.wait_for_samples(n=1, timeout=self.interval * 5)

    current_stats = self.stats()
    if historical_stats:
        stats = aggregate_stats([current_stats] + historical_stats)
    else:
        stats = current_stats

    rec = {}
    # target average CPU usage
    rec["cpu"] = max(1, round(stats["process_cpu_usage"]["mean"]))
    # target maximum memory usage (kB->MB) with a 20% buffer
    rec["memory"] = round_memory(mb=stats["process_memory"]["max"] * 1.2 / 1024)
    # target maximum GPU number of GPUs used
    rec["gpu"] = (
        max(1, round(stats["process_gpu_usage"]["max"]))
        if stats["process_gpu_usage"]["mean"] > 0
        else 0
    )
    # target maximum VRAM usage (MiB) with a 20% buffer
    rec["vram"] = (
        round_memory(mb=stats["process_gpu_vram"]["max"] * 1.2)
        if stats["process_gpu_vram"]["max"] > 0
        else 0
    )
    return rec

recommend_server #

recommend_server(**kwargs)

Recommend the cheapest cloud server matching the recommended resources.

Parameters:

Name Type Description Default
**kwargs

Additional filtering arguments (e.g. vendor_id or compliance_framework_id) to pass to the Spare Cores Keeper API.

 {}

Returns:

Type Description
dict

A dictionary containing the recommended cloud server. Response format is described at https://keeper.sparecores.net/redoc#tag/Query-Resources/operation/search_servers_servers_get.
Source code in resource_tracker/tracker.py 
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
def recommend_server(self, **kwargs) -> dict:
    """Recommend the cheapest cloud server matching the recommended resources.

    Args:
        **kwargs: Additional filtering arguments (e.g. vendor_id or compliance_framework_id) to pass to the Spare Cores Keeper API.

    Returns:
        A dictionary containing the recommended cloud server. Response format is described at <https://keeper.sparecores.net/redoc#tag/Query-Resources/operation/search_servers_servers_get>.
    """
    historical_stats = kwargs.pop("historical_stats", [])
    rec = self.recommend_resources(historical_stats=historical_stats)
    return get_recommended_cloud_servers(**rec, **kwargs, n=1)[0]

SystemTracker #

Track system-wide resource usage.

This class monitors system resources like CPU times and usage, memory usage, GPU and VRAM utilization, disk I/O, and network traffic for the entire system.

Data is collected every interval seconds and written to the stdout or output_file (if provided) as CSV. Currently, the following columns are tracked:

  • timestamp (float): The current timestamp.
  • processes (int): The number of running processes.
  • utime (int): The total user+nice mode CPU time in seconds.
  • stime (int): The total system mode CPU time in seconds.
  • cpu_usage (float): The current CPU usage between 0 and number of CPUs.
  • memory_free (int): The amount of free memory in kB.
  • memory_used (int): The amount of used memory in kB.
  • memory_buffers (int): The amount of memory used for buffers in kB.
  • memory_cached (int): The amount of memory used for caching in kB.
  • memory_active (int): The amount of memory used for active pages in kB.
  • memory_inactive (int): The amount of memory used for inactive pages in kB.
  • disk_read_bytes (int): The total number of bytes read from disk.
  • disk_write_bytes (int): The total number of bytes written to disk.
  • disk_space_total_gb (float): The total disk space in GB.
  • disk_space_used_gb (float): The used disk space in GB.
  • disk_space_free_gb (float): The free disk space in GB.
  • net_recv_bytes (int): The total number of bytes received over network.
  • net_sent_bytes (int): The total number of bytes sent over network.
  • gpu_usage (float): The current GPU utilization between 0 and GPU count.
  • gpu_vram (float): The current GPU memory used in MiB.
  • gpu_utilized (int): The number of GPUs with utilization > 0.

Parameters:

Name Type Description Default
start_time float

Time when to start tracking. Defaults to current time.

 time()
interval float

Sampling interval in seconds. Defaults to 1.

 1
autostart bool

Whether to start tracking immediately. Defaults to True.

 True
output_file str

File to write the output to. Defaults to None, print to stdout.

 None

Methods:

Name Description
__call__

Dummy method to make this class callable.
diff_stats

Calculate stats since last call.
start_tracking

Start an infinite loop tracking system resource usage.
Source code in resource_tracker/tracker.py 
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
class SystemTracker:
    """Track system-wide resource usage.

    This class monitors system resources like CPU times and usage, memory usage,
    GPU and VRAM utilization, disk I/O, and network traffic for the entire system.

    Data is collected every `interval` seconds and written to the stdout or
    `output_file` (if provided) as CSV. Currently, the following columns are
    tracked:

    - timestamp (float): The current timestamp.
    - processes (int): The number of running processes.
    - utime (int): The total user+nice mode CPU time in seconds.
    - stime (int): The total system mode CPU time in seconds.
    - cpu_usage (float): The current CPU usage between 0 and number of CPUs.
    - memory_free (int): The amount of free memory in kB.
    - memory_used (int): The amount of used memory in kB.
    - memory_buffers (int): The amount of memory used for buffers in kB.
    - memory_cached (int): The amount of memory used for caching in kB.
    - memory_active (int): The amount of memory used for active pages in kB.
    - memory_inactive (int): The amount of memory used for inactive pages in kB.
    - disk_read_bytes (int): The total number of bytes read from disk.
    - disk_write_bytes (int): The total number of bytes written to disk.
    - disk_space_total_gb (float): The total disk space in GB.
    - disk_space_used_gb (float): The used disk space in GB.
    - disk_space_free_gb (float): The free disk space in GB.
    - net_recv_bytes (int): The total number of bytes received over network.
    - net_sent_bytes (int): The total number of bytes sent over network.
    - gpu_usage (float): The current GPU utilization between 0 and GPU count.
    - gpu_vram (float): The current GPU memory used in MiB.
    - gpu_utilized (int): The number of GPUs with utilization > 0.

    Args:
        start_time: Time when to start tracking. Defaults to current time.
        interval: Sampling interval in seconds. Defaults to 1.
        autostart: Whether to start tracking immediately. Defaults to True.
        output_file: File to write the output to. Defaults to None, print to stdout.
    """

    def __init__(
        self,
        start_time: float = time(),
        interval: float = 1,
        autostart: bool = True,
        output_file: str = None,
    ):
        _, self.get_system_stats = get_tracker_implementation()

        self.status = "running"
        self.interval = interval
        self.cycle = 0
        self.start_time = start_time

        # dummy data collection so that diffing on the first time does not fail
        self.stats = self.get_system_stats()

        if autostart:
            # wait for the start time to be reached
            if start_time > time():
                sleep(start_time - time())
            # we can now start. 1st interval used to collect baseline
            self.start_tracking(output_file)

    def __call__(self):
        """Dummy method to make this class callable."""
        pass

    def diff_stats(self):
        """Calculate stats since last call."""
        last_stats = self.stats
        self.stats = self.get_system_stats()
        self.cycle += 1

        time_diff = self.stats["timestamp"] - last_stats["timestamp"]

        total_read_bytes = 0
        total_write_bytes = 0
        for disk_name in set(self.stats["disk_stats"]) & set(last_stats["disk_stats"]):
            read_bytes = max(
                0,
                self.stats["disk_stats"][disk_name]["read_bytes"]
                - last_stats["disk_stats"][disk_name]["read_bytes"],
            )
            write_bytes = max(
                0,
                self.stats["disk_stats"][disk_name]["write_bytes"]
                - last_stats["disk_stats"][disk_name]["write_bytes"],
            )
            total_read_bytes += read_bytes
            total_write_bytes += write_bytes

        disk_space_total = 0
        disk_space_used = 0
        disk_space_free = 0
        for disk_space in self.stats["disk_spaces"].values():
            disk_space_total += disk_space["total"]
            disk_space_used += disk_space["used"]
            disk_space_free += disk_space["free"]

        return {
            "timestamp": round(self.stats["timestamp"], 3),
            "processes": self.stats["processes"],
            "utime": max(0, round(self.stats["utime"] - last_stats["utime"], 6)),
            "stime": max(0, round(self.stats["stime"] - last_stats["stime"], 6)),
            "cpu_usage": round(
                max(
                    0,
                    (
                        (self.stats["utime"] + self.stats["stime"])
                        - (last_stats["utime"] + last_stats["stime"])
                    )
                    / time_diff,
                ),
                4,
            ),
            "memory_free": self.stats["memory_free"],
            "memory_used": self.stats["memory_used"],
            "memory_buffers": self.stats["memory_buffers"],
            "memory_cached": self.stats["memory_cached"],
            "memory_active": self.stats["memory_active"],
            "memory_inactive": self.stats["memory_inactive"],
            "disk_read_bytes": total_read_bytes,
            "disk_write_bytes": total_write_bytes,
            "disk_space_total_gb": round(disk_space_total / (1024**3), 2),
            "disk_space_used_gb": round(disk_space_used / (1024**3), 2),
            "disk_space_free_gb": round(disk_space_free / (1024**3), 2),
            "net_recv_bytes": max(
                0, self.stats["net_recv_bytes"] - last_stats["net_recv_bytes"]
            ),
            "net_sent_bytes": max(
                0, self.stats["net_sent_bytes"] - last_stats["net_sent_bytes"]
            ),
            "gpu_usage": self.stats["gpu_usage"],
            "gpu_vram": self.stats["gpu_vram"],
            "gpu_utilized": self.stats["gpu_utilized"],
        }

    def start_tracking(
        self, output_file: Optional[str] = None, print_header: bool = True
    ):
        """Start an infinite loop tracking system resource usage.

        A CSV line is written every `interval` seconds.

        Args:
            output_file: File to write the output to. Defaults to None, printing to stdout.
            print_header: Whether to print the header of the CSV. Defaults to True.
        """
        file_handle = open(output_file, "wb") if output_file else stdout.buffer
        try:
            while True:
                current_stats = self.diff_stats()
                # don't print values yet, we collect data for the 1st baseline
                if self.cycle == 1:
                    if print_header:
                        file_handle.write(
                            render_csv_row(
                                current_stats.keys(), quoting=QUOTE_NONNUMERIC
                            )
                        )
                else:
                    file_handle.write(
                        render_csv_row(current_stats.values(), quoting=QUOTE_NONNUMERIC)
                    )
                if output_file:
                    file_handle.flush()
                # sleep until the next interval
                sleep(max(0, self.start_time + self.interval * self.cycle - time()))
        finally:
            if output_file and not file_handle.closed:
                file_handle.close()

__call__ #

__call__()

Dummy method to make this class callable.

Source code in resource_tracker/tracker.py 
280
281
282
def __call__(self):
    """Dummy method to make this class callable."""
    pass

diff_stats #

diff_stats()

Calculate stats since last call.

Source code in resource_tracker/tracker.py 
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
def diff_stats(self):
    """Calculate stats since last call."""
    last_stats = self.stats
    self.stats = self.get_system_stats()
    self.cycle += 1

    time_diff = self.stats["timestamp"] - last_stats["timestamp"]

    total_read_bytes = 0
    total_write_bytes = 0
    for disk_name in set(self.stats["disk_stats"]) & set(last_stats["disk_stats"]):
        read_bytes = max(
            0,
            self.stats["disk_stats"][disk_name]["read_bytes"]
            - last_stats["disk_stats"][disk_name]["read_bytes"],
        )
        write_bytes = max(
            0,
            self.stats["disk_stats"][disk_name]["write_bytes"]
            - last_stats["disk_stats"][disk_name]["write_bytes"],
        )
        total_read_bytes += read_bytes
        total_write_bytes += write_bytes

    disk_space_total = 0
    disk_space_used = 0
    disk_space_free = 0
    for disk_space in self.stats["disk_spaces"].values():
        disk_space_total += disk_space["total"]
        disk_space_used += disk_space["used"]
        disk_space_free += disk_space["free"]

    return {
        "timestamp": round(self.stats["timestamp"], 3),
        "processes": self.stats["processes"],
        "utime": max(0, round(self.stats["utime"] - last_stats["utime"], 6)),
        "stime": max(0, round(self.stats["stime"] - last_stats["stime"], 6)),
        "cpu_usage": round(
            max(
                0,
                (
                    (self.stats["utime"] + self.stats["stime"])
                    - (last_stats["utime"] + last_stats["stime"])
                )
                / time_diff,
            ),
            4,
        ),
        "memory_free": self.stats["memory_free"],
        "memory_used": self.stats["memory_used"],
        "memory_buffers": self.stats["memory_buffers"],
        "memory_cached": self.stats["memory_cached"],
        "memory_active": self.stats["memory_active"],
        "memory_inactive": self.stats["memory_inactive"],
        "disk_read_bytes": total_read_bytes,
        "disk_write_bytes": total_write_bytes,
        "disk_space_total_gb": round(disk_space_total / (1024**3), 2),
        "disk_space_used_gb": round(disk_space_used / (1024**3), 2),
        "disk_space_free_gb": round(disk_space_free / (1024**3), 2),
        "net_recv_bytes": max(
            0, self.stats["net_recv_bytes"] - last_stats["net_recv_bytes"]
        ),
        "net_sent_bytes": max(
            0, self.stats["net_sent_bytes"] - last_stats["net_sent_bytes"]
        ),
        "gpu_usage": self.stats["gpu_usage"],
        "gpu_vram": self.stats["gpu_vram"],
        "gpu_utilized": self.stats["gpu_utilized"],
    }

start_tracking #

start_tracking(output_file=None, print_header=True)

Start an infinite loop tracking system resource usage.

A CSV line is written every interval seconds.

Parameters:

Name Type Description Default
output_file Optional[str]

File to write the output to. Defaults to None, printing to stdout.

 None
print_header bool

Whether to print the header of the CSV. Defaults to True.

 True
Source code in resource_tracker/tracker.py 
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
def start_tracking(
    self, output_file: Optional[str] = None, print_header: bool = True
):
    """Start an infinite loop tracking system resource usage.

    A CSV line is written every `interval` seconds.

    Args:
        output_file: File to write the output to. Defaults to None, printing to stdout.
        print_header: Whether to print the header of the CSV. Defaults to True.
    """
    file_handle = open(output_file, "wb") if output_file else stdout.buffer
    try:
        while True:
            current_stats = self.diff_stats()
            # don't print values yet, we collect data for the 1st baseline
            if self.cycle == 1:
                if print_header:
                    file_handle.write(
                        render_csv_row(
                            current_stats.keys(), quoting=QUOTE_NONNUMERIC
                        )
                    )
            else:
                file_handle.write(
                    render_csv_row(current_stats.values(), quoting=QUOTE_NONNUMERIC)
                )
            if output_file:
                file_handle.flush()
            # sleep until the next interval
            sleep(max(0, self.start_time + self.interval * self.cycle - time()))
    finally:
        if output_file and not file_handle.closed:
            file_handle.close()