[GH-ISSUE #332] High iodepth exhibits roughly the same read performance as low iodepth by using Usrbio. #204

New issue

Open

opened 2026-10-26 09:47:05 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 09:47:05 +00:00

Owner

Originally created by @carson-yang on GitHub (Aug 27, 2025).
Original GitHub issue: https://github.com/deepseek-ai/3FS/issues/332

I adopt https://github.com/deepseek-ai/3FS/blob/main/src/lib/api/UsrbIo.md#example-4 to read 20GiB file serially (single thread), I use two different configs:
1. NUM_IOS = 640，BLOCK_SIZE = (32 << 20) with just one hf3fs_submit_ios
2. NUM_IOS = 1, BLOCK_SIZE = (32 << 20) with multiple hf3fs_submit_ios
Both give the roughly the same read performance (about 1.5GiB/s), but I expect above first config should have more high degree of parallelism, so it should have more high read throughput performance.（The entire 3fs system uses the default configuration file）

Originally created by @carson-yang on GitHub (Aug 27, 2025). Original GitHub issue: https://github.com/deepseek-ai/3FS/issues/332 I adopt https://github.com/deepseek-ai/3FS/blob/main/src/lib/api/UsrbIo.md#example-4 to read 20GiB file serially (single thread), I use two different configs: 1. NUM_IOS = 640，BLOCK_SIZE = (32 << 20) with just one hf3fs_submit_ios 2. NUM_IOS = 1, BLOCK_SIZE = (32 << 20) with multiple hf3fs_submit_ios Both give the roughly the same read performance (about 1.5GiB/s), but I expect above first config should have more high degree of parallelism, so it should have more high read throughput performance.（The entire 3fs system uses the default configuration file）

bsakaguchi referenced this issue

2026-10-26 09:48:42 +00:00

[PR #204] [CLOSED] FuseOps.cc: delete dead code and unnecessary comments #327