The GPU engines are built per worker per step, and each took its page-locked host buffers with
cudaHostAlloc / cudaHostRegister and gave them back with cudaFreeHost / cudaHostUnregister. Those
calls map and lock pages under the process's memory-map lock, so all workers queue behind each
other: on the 5080 workstation a myoglobin run made ~3400 of them (6-22 thread-s, depending on the
load), 9gdj ~10400 (124 thread-s); on a four-socket host they cost 20-30 ms each.
- CudaHostPtr takes its memory from a pool: a freed buffer is kept in a power-of-two size class for
the next engine, up to 512 MiB in all; a buffer above 64 MiB is never kept. Rugnux::Run hands the
pool back at its end.
- PinnedVector (std::vector with the pool's allocator) replaces the registered vectors of
AdaptiveSpotFinderGPU and ImagePreprocessorGPU; the Bragg predictors download into a pooled
pinned buffer and copy into their output, instead of registering it.
- BSLZ4DecoderGPU copies the compressed chunk into a pooled pinned buffer before the upload, so it
is one DMA instead of the driver's pageable staging (9gdj: 163 GB pageable H2D, 52 s of copy
time, -> 147 GB pinned, 13 s).
Measured with nsys on the workstation: myob 3398 -> 1396 calls (674 of them the release at the
end), cytc 2986 -> 1270, 9gdj 10428 -> 1748 (124 -> 3.7 thread-s). Output identical (p.hkl md5 on
myob, cytc, 9gdj; CI tier battery).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EizBKhTqrAYago9KJG3dA3
leonarski_fandClaude Opus 5.5
09c41d067d