[LU-3970] Add procfs interface for clearing lustre caches in parallel

• Type: Improvement
• Resolution: Won't Fix
• Priority: Minor
• Labels: patch performance

Epic/Theme

Performance

• Assignee: Jian Yu
• Reporter: Ryan Haasken

Description

Cray experienced slowness clearing the lustre caches after job termination. This slowness is a result of clearing the caches for each namespace in series. Lustre should provide a high-level procfs interface which clears lustre caches across namespaces in parallel and returns when all caches are cleared.

The interface should be at /proc/fs/lustre/ldlm/drop_caches, and when written to, it must clear all lustre caches in parallel and return when finished.

It must have the same effect as the following, but in parallel:

for LRU in $(ls /proc/fs/lustre/ldlm/namespaces/*osc*/lru_size); do
    echo clear > $LRU
done

for LRU in $(ls /proc/fs/lustre/ldlm/namespaces/*mdc*/lru_size); do
    echo clear > $LRU
done

Issue Links

• is related to: LU-5134 Add option to lctl set_param for setting parameters in parallel

Comments

[Peter Jones]

ok Cory

[Cory Spitz]

http://review.whamcloud.com/#/c/7783 is abandoned in favor of the approach taken with LU-5134.

This bug ought to be closed now. Jian, can you please make it so?

[Ryan Haasken]

Can somebody please mark this bug related to LU-5134?

LU-5134 will resolve this issue by allowing lctl set_param to spawn threads in user space when setting lru_size=clear.

[Ryan Haasken]

I have submitted a patch to Gerrit: http://review.whamcloud.com/#/c/7783/

[Ryan Haasken]

That would work as well. However, Cray's ALPS (Application Level Placement Scheduler) team requested that Lustre provide a higher level interface which clears all the caches in parallel. Do you feel that this is an appropriate enhancement? I've already tested a patch, but I am still learning how to submit it to Gerrit for review.

[Oleg Drokin]

I wonder if why doing hte echos in parallel won't work?

[Ryan Haasken]

I am working on uploading a patch to Gerrit.