When running many trials and sessions some will get slowed down with low GPU/CPU usage. This happens across GPU and CPU, with and without Ray, with new and old version of PyTorch, and even with careful garbage collection. The slowdown is to 5 FPS for cartpole and 50 FPS for Atari with venv.
After debugging, the cause seems to be XVFB wrapper for Linux. Suspicion is that the fake I/O gets overcrowded and slowed for some processes and they remain so.
removed XVFB and the normal FPS is restored.
Enable concurrent Ray runs
Multiple ray run is useful for running many searches/benchmarks.
allow multiple Ray runs by assigning the Redis port automatically by PID
Fix XVFB slowdown
When running many trials and sessions some will get slowed down with low GPU/CPU usage. This happens across GPU and CPU, with and without Ray, with new and old version of PyTorch, and even with careful garbage collection. The slowdown is to 5 FPS for cartpole and 50 FPS for Atari with venv.
After debugging, the cause seems to be XVFB wrapper for Linux. Suspicion is that the fake I/O gets overcrowded and slowed for some processes and they remain so.
Enable concurrent Ray runs
Multiple ray run is useful for running many searches/benchmarks.
More search specs