hmacdope / fah-gpu-species

fah-gpu-analysis (EXPERIMENTAL)

DISCLAIMER: the GPU benchmark data in this repo is experimental and not final. For now, the code and data presented here is intended as an example of parsing performance data from fah-core logs, and of the types of analysis that might be useful to automate intelligent assignment of work units to specific devices.

Analysis to understand GPU performance on Folding@home benchmarking tasks. The eventual goal is to create a data-driven automatic assignment of GPU species.

GPU species is an 8-bit unsigned integer (range 0 to 255) assigned to each distinct device known to F@H. It is used to constrain which GPUs can run work for a given project. For example, a project might have the setting NVIDIAGPUSpecies >= 3, which means that work units will only be assigned to NVIDIA GPUs with a GPUSpecies of 3 or higher.

See notebooks for example analyses.

Installation

$ python setup.py install

Analyzing science.log files

While it runs, the F@H core writes a log file called science.log that contains, among other things, GPU device information and timing data useful for assessing the relative performance of GPUs. The command parse_fah_log can be used to extract useful information from these files. It can either be used as a library, deserializing to Python dataclass instances for further analysis, or as a CLI tool producing JSON.

1. To download log files from work server, set FAH_HOST, SERVER_ID, and PROJ_ID in your environment and run

   rsync --include '*/' --include '*/science.log' --exclude '*' --recursive $FAH_HOST:path/to/SVR${SERVER_ID}/PROJ${PROJ_ID} ./

2. Because the log files are generated by clients running different OSes, they generally have a mixture of encodings and line delimiters. Use dos2unix or similar to convert to Unix format, e.g.

   find PROJ${PROJ_ID}/ -name science.log | parallel --progress dos2unix {}

3. Follow the usage example below to parse individual log files. See the notebooks folder for example analyses.

Usage example

$ parse_fah_log /path/to/science.log | jq '.fah_core_log'

{
  "version": {
    "major": 0,
    "minor": 0,
    "patch": 11
  },
  "platforms": [
    {
      "info": {
        "profile": "FULL_PROFILE",
        "version": "OpenCL 1.2 CUDA 11.0.126",
        "name": "NVIDIA CUDA",
        "vendor": "NVIDIA Corporation"
      },
      "devices": [
        {
          "name": "P106-090",
          "vendor": "NVIDIA Corporation",
          "version": "OpenCL 1.2 CUDA"
        },
        {
          "name": "P106-090",
          "vendor": "NVIDIA Corporation",
          "version": "OpenCL 1.2 CUDA"
        },
        {
          "name": "P106-090",
          "vendor": "NVIDIA Corporation",
          "version": "OpenCL 1.2 CUDA"
        },
        {
          "name": "P106-090",
          "vendor": "NVIDIA Corporation",
          "version": "OpenCL 1.2 CUDA"
        }
      ]
    },
    {
      "info": {
        "profile": "FULL_PROFILE",
        "version": "OpenCL 2.0 AMD-APP (1800.11)",
        "name": "AMD Accelerated Parallel Processing",
        "vendor": "Advanced Micro Devices, Inc."
      },
      "devices": [
        {
          "name": "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz",
          "vendor": "GenuineIntel",
          "version": "OpenCL 1.2 AMD-APP (1800.11)"
        }
      ]
    },
    {
      "info": {
        "profile": "FULL_PROFILE",
        "version": "OpenCL 1.2 ",
        "name": "Intel(R) OpenCL",
        "vendor": "Intel(R) Corporation"
      },
      "devices": [
        {
          "name": "Intel(R) Core(TM) i7-3770 CPU @ 3.40GHz",
          "vendor": "Intel(R) Corporation",
          "version": "OpenCL 1.2 (Build 76427)"
        },
        {
          "name": "Intel(R) HD Graphics 4000",
          "vendor": "Intel(R) Corporation",
          "version": "OpenCL 1.2 "
        }
      ]
    }
  ],
  "checkpoint_perfs_ns_day": [
    38.96129152,
    38.96129152,
    38.86139077,
    38.96129152
  ],
  "average_perf_ns_day": 38.9613
}