[LU-17122] reduce default osd-ldiskfs.*.readcache_max_io_mb and writethrough_max_io_mb

Fri, 15 Sep 2023 23:08:32 +0000

We currently use osd-ldiskfs.*.writethrough_max_io_mb=8 and "...readcache_max_io_mb=8" as the default thresholds to switch from page cached to direct write/read RPCs on the OSS. This is mostly important for HDD OSTs, since non-rotational OSTs disable write and read cache by default.

However, I'm not aware of any performance analysis of these thresholds, whether any testing has been done to lower these thresholds, or somehow align the sizes to the underlying RAID geometry.

sihara, have you ever tested/tuned these parameters to see what effect they have on IO performance? Ideally we would have them as low as possible to reduce server memory pressure, while keeping good IO performance?

I don't think this is a critical issue, but if you could include it in one of your HDD test runs to see what effect it has, and whether we are losing performance that could be easily improved by changing the default values.

Whamcloud Community JIRA

[LU-17122] reduce default osd-ldiskfs.*.readcache_max_io_mb and writethrough_max_io_mb