[LU-18556] Rework HSM to use llog cookie for a record modification

• Type: Improvement
• Resolution: Unresolved
• Priority: Major

Epic/Theme

hsm

• Severity: 3
• Assignee: Alexander Boyko
• Reporter: Alexander Boyko

Description

llog_write could use llog cookie for a single record modification. In this case there is no need to process all records before an update of the record. HSM regular updates record, and cancels it after timeout only, so there could be many records at the same time. And this introduce a bottleneck for HSM.

llog subsystem requires some additional functions to do it. Also HSM need to store a whole llog record for a modification, and this leads to a different structures in memory for a HSM.
Now, Lustre operates and stores hsm_action_item, however llog_agent_req_rec include hai already. So this change is some trade of for a memory vs additional processing.

Comments

[Alexander Boyko]

We have a testing results for this patch at perf cluster, many thanks to Nikitas.
Test 1

• Adding 1000.000 archive requests
• start copy tool to process requests with no operation (get req and reply, to avoid copytool processing time, to measure only Lustre)
• measuring the time needed for the copytool to handle the first 1,000,000 requests

Test 2
Similar to test 1, but also queueing an additional 1,000,000 requests in the MDT, immediately after start of processing, measuring the time needed for the copytool to handle the first 1,000,000 requests.

	base	LU-18556 patch
Test 1 seconds	572	187
Test 2 seconds	558	392

The patch should offer a significant performance benefit compared to current 2.16

[Alex Zhuravlev]

the idea of updating a llog record through llog processing - has much overhead.

there is no need to go through the llog looking for a specific record – if you have something like an index.

[Alexander Boyko]

I'll take a look. But llog improvement would not help much to HSM, because the idea of updating a llog record through llog processing - has much overhead. With massive HSM archive, release etc., it is disaster. Llog was not designed for such using.
We had an attempt to replace the HSM actions queue llog with index, in an effort to address a performance issues. But the massive testing showed it became worse then llog.

[Alex Zhuravlev]

HSM uses local llog, so it is not related to DNE3 problem.

that approach solves "local" problems as well - removes need for expensive serialization.

[Alexander Boyko]

HSM uses local llog, so it is not related to DNE3 problem. Also I've pushed an improvement for changelog+mdtest https://review.whamcloud.com/c/fs/lustre-release/+/56342

[Alex Zhuravlev]

I'd suggest to look at LU-7426

[Gerrit Updater]

"Alexander Boyko <[e-mail]>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/57428
Subject: LU-18556 hsm: use direct modifying llog record
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 780d31e9b83896d18aa7cda52ec59c2a4a035f66