Blobovnizca PUT performance degradation #602

Closed
opened 2023-08-11 12:17:21 +00:00 by dstepanov-yadro · 7 comments

Current Behavior

Blobovnizca works well in case of low load (around 10-15RPS). But if there are around 30 RPS, then response time in 0.95 quantile goes up to 10 seconds.

See attachment: after RPS increasing at 17:45 latency raises up to 10 seconds

Expected Behavior

I think it must be some linear degradation or something like this. 10 seconds is too much.

Your Environment

MetalCore
12 HDD + 2 SSD
Payload 8KB

image

## Current Behavior Blobovnizca works well in case of low load (around 10-15RPS). But if there are around 30 RPS, then response time in 0.95 quantile goes up to 10 seconds. See attachment: after RPS increasing at 17:45 latency raises up to 10 seconds ## Expected Behavior I think it must be some linear degradation or something like this. 10 seconds is too much. ## Your Environment MetalCore 12 HDD + 2 SSD Payload 8KB ![image](/attachments/035c5438-6749-4d3d-b878-a60df748278a)
397 KiB
dstepanov-yadro added the
bug
triage
frostfs-node
labels 2023-08-11 12:17:44 +00:00
Collaborator

Does it happen with any payload size as well?

Does it happen with any payload size as well?
Poster
Collaborator

Does it happen with any payload size as well?

Haven't tested all possible values.

> Does it happen with any payload size as well? Haven't tested all possible values.

Does it have any dependency on blobovniczas depth/width?

Does it have any dependency on blobovniczas depth/width?

30 RPS per blobovnicza tree, NOT per node, right?

30 RPS per blobovnicza tree, NOT per node, right?
Poster
Collaborator

30 RPS per blobovnicza tree, NOT per node, right?

Right, per shard.

Does it have any dependency on blobovniczas depth/width?

Haven't tested again.

I think this bug requires some research & development, not only development.

> 30 RPS per blobovnicza tree, NOT per node, right? Right, per shard. > Does it have any dependency on blobovniczas depth/width? Haven't tested again. I think this bug requires some research & development, not only development.

Had a thought that this is related to opened_cache_size (so dbs are closed an opened all the time), but it must've been 1000 in this test.

Had a thought that this is related to `opened_cache_size` (so dbs are closed an opened all the time), but it must've been 1000 in this test.
dstepanov-yadro self-assigned this 2023-08-11 13:42:45 +00:00
fyrchik added this to the v0.37.0 milestone 2023-08-14 07:11:08 +00:00
Poster
Collaborator

Degradation occurs if a large number of databases are open.
As a solution, the ability to specify the number of databases on leaf nodes has been added.
The calculation of the database size has also been fixed.
It is recommended not to use databases larger than 40-50 MB.

Degradation occurs if a large number of databases are open. As a solution, the ability to specify the number of databases on leaf nodes has been added. The calculation of the database size has also been fixed. It is recommended not to use databases larger than 40-50 MB.
Sign in to join this conversation.
No Milestone
No Assignees
3 Participants
Notifications
Due Date
The due date is invalid or out of range. Please use the format 'yyyy-mm-dd'.

No due date set.

Dependencies

No dependencies set.

Reference: TrueCloudLab/frostfs-node#602
There is no content yet.