Skip to content

Increase batch size for Allen and modified host memory to support this

Da Yu Tou requested to merge dtou_increase_batch_size_and_host_memory into 2024-patches

Allen runs faster with larger batch size. This MR changes the Allen configurations to use batch sizes of 1000 and host memory of 400Mb.

Now that Allen!1679 (merged) is merged, the change to host memory is not urgent but nice to have since we might be increasing the amount of algorithms in Allen (commented by @ahennequ below). Previously, host runs out of memory due to host-side prefix sum memory pressure but Allen!1679 (merged) will move all prefix sum operations onto the device which makes the change to host memory unnecessary.

Improvement With events-per-slice

The throughput of hlt1_pp_matching_no_ut with different events-per-slice is shown below. This MR will increase it from 500 to 1000. Larger output sizes will also relieve some IO pressure when writing out events from Allen. image

Edited by Da Yu Tou

Merge request reports