SherlockBench: Windows CPU & Memory Workload Telemetry

Release 0.5.0 contains 7,380 measured observations, 123 five-minute experiments, 8 reported Windows devices and 11 device/session pairs. This release adds three submitted collection batches (pc-03/s01, pc-04/s01 and pc-08/s01): 36 runs were reviewed, 35 were accepted into the primary table and one was retained only in raw/excluded.

The current primary device set includes pc-01 through pc-08. Device IDs are pseudonymous collector labels, not independently verified hardware identities.

Kaggle · Hugging Face · Source project

Workloads are generated programmatically; measurements are real. No primary observations were duplicated, imputed or resampled. The 123 accepted runs represent 10 hours 15 minutes of nominal experiment windows, not continuous CPU measurement or a continuous recording.

What changed in 0.5.0

• Added pc-03/s01: 12 accepted runs / 720 rows.
• Added pc-04/s01: 12 accepted runs / 720 rows.
• Added pc-08/s01: 11 accepted runs / 660 rows.
• Retained one pc-08/s01 mixed run outside primary because it has 59 rows and a 10.391-second inter-sample gap, above the 10-second release threshold.
• Canonicalized derived session IDs to the compact sNN form. For example, session-002 is now exported as s02.
• Kept historical collector-emitted labels unchanged inside raw/ metadata so the source provenance remains auditable. telemetry.csv, runs.csv, reports and documentation use the canonical names.

Scenarios

Scenario	Runs	Rows	Injected workload
normal	31	1,860	No injected load; background applications uncontrolled
cpu_load	31	1,860	Scheduled CPU workers
memory_growth	31	1,860	Gradual RAM allocation
mixed	30	1,800	CPU and RAM workloads together

Each accepted run has 12 baseline, 36 active and 12 recovery observations. Requested phase lengths are 60/180/60 seconds with a five-second sampling interval. CPU readings cover approximately one second of each interval. Active phases include workload ramps.

Device and session coverage

Device	Session	Runs	Rows	Logical CPUs	RAM bytes
pc-01	s02	12	720	16	29,797,376,000
pc-01	s03	12	720	16	29,797,376,000
pc-02	s01	4	240	28	34,163,507,200
pc-02	s02	12	720	28	34,163,507,200
pc-03	s01	12	720	12	16,951,447,552
pc-04	s01	12	720	12	17,002,815,488
pc-05	s03	12	720	16	29,854,978,048
pc-06	s01	12	720	12	34,194,051,072
pc-06	s02	12	720	12	34,194,051,072
pc-07	s01	12	720	16	17,102,315,520
pc-08	s01	11	660	12	34,189,307,904

pc-01/s02 is the canonical exported name for the historical collector label pc-01/session-002. Historical session-001 and smoke-002 labels are likewise mapped to s01 and s02 in derived inventory tables. Missing session numbers are not reconstructed.

For CPU and mixed scenarios, pc-02/s01 requested one worker at duty 0.5 and a 128 MiB memory target where applicable. Other current primary batches use two workers at duty 0.5 and 256 MiB where applicable. Duty is per worker, not whole-PC CPU utilization; memory allocation targets are requested values, not independently measured achieved allocations.

Files

• telemetry.csv: primary joined table, 7,380 rows. Start here.
• runs.csv: 131-run inventory across all quality categories.
• data_dictionary.csv: primary table fields, units and roles.
• quality_report.json: validation results, exclusions and inclusion policy.
• SherlockBench_EDA.ipynb: executable exploratory analysis, no model-score claims.
• figures/workload_timelines.png: primary timelines by scenario and device.
• raw/main/: 123 accepted source metadata/CSV pairs.
• raw/batches/: nine submitted batch manifests retained for provenance.
• raw/excluded/: two excluded runs (106 rows total).
• raw/smoke/: four installation checks with 28 rows.
• raw/legacy_pilot/: two earlier recordings with 179 rows and a different schema.
• code/rebuild_tables.py: validates reviewed raw/main and rebuilds canonical tables.
• code/scenario_batch.py: historical recorder source retained from earlier releases.
• checksums.sha256: release file checksums.

All raw categories contain 7,693 observations, of which 7,380 are primary. The joined primary table repeats raw/main observations for convenience: do not append raw files to it. Other categories remain outside the primary table.

Labels: what they mean

scenario describes the entire experiment. It does not mean a load was active in every row. phase identifies baseline, active or recovery. scheduled_load is derived: scenario during active workload phases, otherwise none. It describes the schedule, not a measured fault. Recovery can retain effects after the injected load stops.

The scheduled-load counts are 4,068 none, 1,116 cpu_load, 1,116 memory_growth and 1,080 mixed. High CPU is not necessarily a fault; memory growth is not proof of a leak. This release is intended for telemetry exploration and prototype workload recognition, not validated root-cause diagnosis, malware detection or production anomaly-detection benchmarks.

Collection and provenance

Release 0.5.0 builds on 0.4.0 and adds pc-03/s01, pc-04/s01 and pc-08/s01. New collection uses Python and psutil 7.2.2 on Windows, schema sherlockbench-system-0.2, recorder 2.1 and collector app 0.1.0. Original run IDs and measured sample values are preserved.

Canonical session names are a presentation-layer normalization in derived exports. Historical raw metadata remains collector-authentic, so a raw file may still contain session-002 even though its corresponding row in telemetry.csv/runs.csv uses s02.

Primary exports contain no process names, usernames, file contents or absolute timestamps. Hardware characteristics and general activity notes in legacy metadata remain included. Data minimization does not guarantee complete anonymity.

Quality review

All 123 primary runs have matching row counts and IDs, sequential sample indices, finite numeric values, valid phase bounds and successful worker exits. A reviewed run is excluded from primary if it does not have 60 observations or if any consecutive sample-start gap exceeds twice the requested interval. This includes gaps across phase boundaries.

Two runs are retained in raw/excluded:

• 2d2bf859ca5343a0aac7f21f416df735: 47 rows and an approximately 90-second maximum gap (historical exclusion from an earlier release).
• 1c713deb2a224d72b29ec7fdcef06ffb: 59 rows and a 10.391-second maximum gap in the new pc-08/s01 submission. No missing row was manufactured or imputed.

All 12 pc-03/s01 runs, all 12 pc-04/s01 runs and 11 of 12 pc-08/s01 runs pass the same primary checks. Background spikes are retained as measured rather than relabelled.

Evaluation limitations

Eight reported machines are still a small exploratory sample. Sessions per device are uneven, device identities are self-assigned, background activity is uncontrolled and hardware/workload strength can be confounded. No benchmark accuracy or broad cross-device generalization is claimed.

There is no official train/test split. The Hugging Face train split is a storage container for all primary rows, not a leakage-safe evaluation split. Keep each run intact; hold out sessions or devices for stronger exploratory evaluations. Never randomly split neighboring rows. Do not use scenario, phase, scheduled_load, run/session/machine IDs, sample indices or elapsed schedule time as classification features. Fit transformations on training data only. Completed-baseline subtraction in the notebook is retrospective and is not online inference.

Quick start

import pandas as pd
samples = pd.read_csv('telemetry.csv')
print(samples.groupby(['machine_id', 'session_id']).size())

from datasets import load_dataset
samples = load_dataset('fireaideveloper/SherlockBench', split='train')

Notebook dependencies: pandas, numpy, matplotlib. No load generation or network access is required for local analysis. Run python code/rebuild_tables.py without -O to validate and regenerate tables, then rerun the notebook after data changes.

License

Creative Commons Attribution 4.0 International (CC BY 4.0). https://creativecommons.org/licenses/by/4.0/

Attribution: Kirill Ivanov (fireaideveloper), SherlockBench, release 0.5.0.