leonarski_f and Claude Opus 5.5 4b1d7feebd
GPU threads: block instead of spin, and stay on the GPU's NUMA node ...
- set_gpu_blocking_sync(): every device is put in
cudaDeviceScheduleBlockingSync before its context exists, so a host thread
waiting on the GPU sleeps instead of spinning on a core. On a 16M rotation
run a fifth of all CPU time was that spinning; wall time unchanged within
noise. Called first thing in rugnux.
- enable_gpu_numa_binding(): from then on pin_gpu() (and the new
pin_gpu(dev), used by the first-pass spot workers that take a card by
index) also keeps the thread on the CPUs of the NUMA node the card hangs
off. The node and its CPUs come from /sys (no libnuma), intersected with
the process's own mask; Linux only, and nothing happens on a machine with a
single node. rugnux turns it on; the broker does not.
- A thread inherits its creator's affinity, so the shared ParallelFor pool
would run every later pass on one socket if a pinned worker created it:
its threads now reset to the mask the process started with
(common/ThreadAffinity).
Byte-identical output. The NUMA part is a no-op on the single-node test box
and still has to be measured on a two-socket machine.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com >
Claude-Session: https://claude.ai/code/session_01D1G8gJVAy6gp1K5Dz3NE5C
2026-09-26 19:30:01 +02:00
2026-06-23 20:29:49 +02:00
2026-06-23 20:29:49 +02:00
2026-04-29 09:50:50 +02:00
2026-08-25 08:21:39 +02:00
2026-08-25 08:21:39 +02:00
2026-08-13 17:03:10 +02:00
2026-08-13 17:03:10 +02:00
2026-08-13 17:03:10 +02:00
2026-08-13 17:03:10 +02:00
2026-09-09 07:25:13 +02:00
2026-09-09 07:25:13 +02:00
2026-03-05 22:13:12 +01:00
2026-06-08 08:30:35 +02:00
2026-09-26 19:30:01 +02:00
2026-08-13 17:03:10 +02:00
2026-08-13 17:03:10 +02:00
2025-05-28 18:49:27 +02:00
2026-08-25 08:21:39 +02:00
2026-09-02 21:17:31 +02:00
2026-09-02 21:17:31 +02:00
2026-09-22 06:48:37 +02:00
2026-09-22 06:48:37 +02:00
2026-09-02 21:17:31 +02:00
2026-09-02 21:17:31 +02:00
2026-09-26 19:30:01 +02:00
2026-09-26 19:30:01 +02:00
2026-09-26 19:30:01 +02:00
2026-04-29 09:50:50 +02:00
2026-06-08 08:30:35 +02:00
2026-09-09 07:25:13 +02:00
2026-09-22 06:48:37 +02:00
2026-09-09 07:25:13 +02:00
2026-06-08 08:30:35 +02:00
2026-04-16 11:59:59 +02:00
2026-06-08 08:30:35 +02:00
2026-04-16 11:59:59 +02:00
2026-06-08 08:30:35 +02:00
2025-11-19 09:40:50 +01:00
2026-06-08 08:30:35 +02:00
2026-09-02 21:17:31 +02:00
2026-09-02 21:17:31 +02:00
2026-04-29 09:50:50 +02:00
2026-06-08 08:30:35 +02:00
2026-09-23 21:04:16 +02:00
2026-09-23 21:04:16 +02:00
2026-08-25 08:21:39 +02:00
2026-08-25 08:21:39 +02:00
2026-09-23 21:04:16 +02:00
2026-09-02 21:17:31 +02:00
2026-09-09 07:25:13 +02:00
2026-09-02 21:17:31 +02:00
2026-08-13 17:03:10 +02:00
2026-08-13 17:03:10 +02:00
2026-09-22 06:48:37 +02:00
2026-07-19 09:39:28 +02:00
2026-06-08 08:30:35 +02:00
2026-09-15 17:09:31 +02:00
2026-09-15 17:09:31 +02:00
2026-09-15 17:09:31 +02:00
2026-08-25 08:21:39 +02:00
2026-08-25 08:21:39 +02:00
2026-09-02 21:17:31 +02:00
2026-06-08 08:30:35 +02:00
2026-09-22 06:48:37 +02:00
2026-02-01 13:29:33 +01:00
2026-06-23 20:29:49 +02:00
2026-06-08 08:30:35 +02:00
2026-04-29 09:50:50 +02:00
2026-06-08 08:30:35 +02:00
2026-09-22 06:48:37 +02:00
2026-09-22 06:48:37 +02:00
2024-12-02 21:17:14 +01:00
2026-06-08 08:30:35 +02:00
2026-09-02 21:17:31 +02:00
2026-06-23 20:29:49 +02:00
2026-09-23 21:04:16 +02:00
2026-09-09 07:25:13 +02:00
2026-09-09 07:25:13 +02:00
2026-08-27 22:16:54 +02:00
2025-09-22 19:49:39 +02:00
2026-06-08 08:30:35 +02:00
2026-04-16 11:59:59 +02:00
2026-06-08 08:30:35 +02:00
2026-04-16 11:59:59 +02:00
2026-06-08 08:30:35 +02:00
2026-09-02 21:17:31 +02:00
2026-06-08 08:30:35 +02:00
2026-06-23 20:29:49 +02:00
2026-06-08 08:30:35 +02:00
2026-09-26 19:30:01 +02:00
2026-09-23 20:42:46 +02:00
2026-09-23 20:42:46 +02:00
2026-09-09 07:25:13 +02:00
2026-06-08 08:30:35 +02:00
2026-06-08 08:30:35 +02:00
2026-06-08 08:30:35 +02:00
2026-09-24 07:26:42 +02:00
2026-08-13 17:03:10 +02:00
2026-09-02 21:17:31 +02:00
2026-09-16 18:17:46 +02:00
2026-06-23 20:29:49 +02:00
2026-06-23 20:29:49 +02:00
2026-06-23 20:29:49 +02:00
2026-09-16 18:17:46 +02:00
2026-06-23 20:29:49 +02:00
2024-12-08 13:26:13 +01:00
2026-06-08 08:30:35 +02:00
2026-06-23 20:29:49 +02:00
2026-06-23 20:29:49 +02:00
2026-06-23 20:29:49 +02:00
2026-06-08 08:30:35 +02:00
2026-09-22 06:48:37 +02:00
2026-09-22 06:48:37 +02:00
2025-06-15 17:46:46 +02:00
2026-09-09 07:25:13 +02:00
2026-09-09 07:25:13 +02:00
2026-06-08 08:30:35 +02:00
2025-09-08 20:28:59 +02:00
2026-06-08 08:30:35 +02:00
2026-04-16 11:59:59 +02:00
2026-06-08 08:30:35 +02:00
2026-08-25 08:21:39 +02:00
2026-09-26 19:30:01 +02:00
2026-09-26 19:30:01 +02:00
2026-09-22 06:48:37 +02:00
2026-06-23 20:29:49 +02:00
2025-12-12 21:24:20 +01:00
2026-06-08 08:30:35 +02:00
2026-05-28 18:48:35 +02:00
2026-05-28 18:48:35 +02:00
2025-09-21 19:27:51 +02:00
2026-06-08 08:30:35 +02:00
2026-08-13 17:03:10 +02:00
2026-06-08 08:30:35 +02:00
2026-04-16 11:59:59 +02:00
2026-09-22 06:48:37 +02:00
leonarski_fandClaude Opus 5.5
4b1d7feebd