Skip to content

CHANGELOG.md

Latest commit

 

History

History
123 lines (107 loc) · 5.91 KB

CHANGELOG.md

File metadata and controls

123 lines (107 loc) · 5.91 KB

Changelog

All notable changes to gb10-gups are documented here.

Added

Initial three-file suite

  • cpu_gups.c — CPU-only GUPS baseline, plain malloc, auto-sizes table from live /proc/meminfo MemAvailable.
  • cuda_gups.cu — GPU-only GUPS baseline, cudaMalloc, explicit table size argument (CUDA memory queries unreliable on GB10).
  • managed_gups.cu — CPU+GPU concurrent GUPS, cudaMallocManaged, configurable CPU/GPU update split.
  • LFSR random generator, starts() seed-jump, and XOR read-modify-write update logic ported from HPCC RandomAccess / OpenSHMEM GUPS (Copyright 2011-2015 University of Houston System and UT-Battelle, LLC, BSD-licensed). SHMEM multi-process partitioning machinery intentionally not reused — does not apply to single-chip unified memory.

License compliance fix

  • Added the verbatim BSD 3-Clause notice (full text, not paraphrased) to all three files' headers, per that license's own redistribution terms. Original provenance description alone was not sufficient.

Runtime platform classification

  • Ported query_um_paradigm() byte-for-byte from cuda-unified-memory-analyzer's um_analyzer_v8.cu, verified against the original source via direct diff (zero logic differences). Both .cu files now self-report FULL_EXPLICIT / FULL_HARDWARE_COHERENT / FULL_SOFTWARE_COHERENT / LIMITED / UNKNOWN at runtime, correctly identifying whichever GPU architecture they are actually run on rather than assuming the build target.

Thermal and power sampling

  • Added background-thread NVML sampling (ThermalSampler, thermal_sample_thread) to all three binaries, sampling every 0.5s for the full run duration — catches real climb on long runs, correctly shows nothing meaningful on sub-second runs.
  • Added read_power_w() implementing the spbm_hwmon-first, NVML-fallback, explicit-"unavailable"-third chain, ported from the real sparkview/power.py. Power source is recorded and reported so readers know which sensor produced a given number.
  • Added average power tracking alongside peak (matching the gb10-kernel-probe forum baseline's avg+peak reporting pattern, not peak alone).

NVML throttle-reason bitmask tracking

  • Added nvmlDeviceGetCurrentClocksThrottleReasons reading and accumulation (throttle_flags_seen), with bitmask constants ported from spark-gpu-throttle-check.py.
  • Fixed the misleading generic "possible throttle" message: the output now distinguishes a real driver-reported throttle flag from a clock drop with no flag set. On Pascal, confirmed the latter case in practice — see Fixed section below.

Structured JSON results logging

  • Added write_results_json() to all three binaries — writes results/<binary>_log2_<n>_<timestamp>.json per run, including the full per-sample thermal/power/throttle time series (not just summary stats), table size, GUP/s, and UM paradigm classification.

Progress display

  • Added live progress reporting to cpu_gups.c's update loop and managed_gups.cu's CPU thread (the only loops instrumentable from the host side — a running GPU kernel cannot report progress mid-flight). cuda_gups.cu uses cudaEventQuery polling instead of a blocking sync, printing elapsed time while the async kernel runs.
  • managed_gups.cu's progress line additionally shows live temp/clock/power read from the shared ThermalSampler, so there is real machine state to look at during multi-minute runs, not just a percentage.

Fixed

  • Forward-declaration ordering bug: wall_seconds() was used inside thermal_sample_thread before its definition later in managed_gups.cu (pre-existing in that file for the CPU thread). Caused a real compile failure (identifier "wall_seconds" is undefined), confirmed and fixed with a forward declaration.
  • ETA calculation noise: dividing by a near-zero pct in the first few progress samples produced wildly inflated ETA estimates (observed: "ETA 29046s" / ~8 hours at 0.2% complete on a run that actually completed in ~250-300s). Fixed by suppressing the ETA display until at least 2% complete and 5 seconds elapsed.
  • Premature min_clock_mhz display: the live progress line was reading min_clock_mhz before it had been set past its 0xFFFFFFFF sentinel, printing a value that looked real but was not yet meaningful. Fixed by gating the display on an explicit "has this been set" check; shows "MHz(measuring)" until a real minimum exists.
  • Stray trailing text on completion line: the "100.0% complete | done" message did not fully overwrite the longest possible prior progress line, leaving fragments like "peak)" visible after completion. Fixed by padding the completion message with sufficient trailing whitespace.
  • Mislabeled "GB10 GUPS" banner on Pascal hardware: initial versions of cuda_gups.cu / managed_gups.cu printed a static "GB10 GUPS" banner regardless of actual hardware. Fixed by adding the query_um_paradigm() port (see Added) so the banner and classification reflect the real detected hardware.

Verified

All three binaries compiled (gcc -Wall zero warnings; nvcc -arch=sm_61 zero errors) and ran to completion repeatedly on real hardware (NVIDIA GTX 1080, SM 6.1, driver 570.195.03):

  • cpu_gups: multiple runs at 2^22, 2^27, 2^29 (auto-sized).
  • cuda_gups: multiple runs at 2^24, 2^29.
  • managed_gups: multiple runs at 2^24, 2^29 (50/50 CPU/GPU split).

Confirmed working as designed: throttle-flag discrimination correctly distinguished Pascal's idle-clock-state cycling (observed: clock dropping to 139 MHz during long managed_gups runs, zero NVML throttle flags set throughout) from a real hardware throttle event, cross-checked against an independent clean PASS from spark-gpu-throttle-check.py run minutes apart on the same card.

No run has been performed on GB10 hardware. All numbers produced to date are from the Pascal control case described above.