perf: performance regression on RisingWave with 0.11

[MrCroxx]

There is an obvious regression on nexmark Q7 and Q014 for RisingWave with foyer 0.11 rather than with foyer 0.10.

Based on the flamegraph, the regression comes from 3 stacks:

Which are:

and

and

Let's see why.

Full flamegraphs:

Grafana:

010: https://grafana.test.risingwave-cloud.xyz/d/EpkBw5W4k/risingwave-dev-dashboard?orgId=1&var-datasource=cdtasocg64074c&var-namespace=xx-foyer-010-1&var-instance=benchmark-risingwave&var-pod=All&var-component=All&var-table=All&from=1726042331594&to=1726086587137
011: https://grafana.test.risingwave-cloud.xyz/d/EpkBw5W4k/risingwave-dev-dashboard?orgId=1&var-datasource=cdtasocg64074c&var-namespace=xx-foyer-011-1&var-instance=benchmark-risingwave&var-pod=All&var-component=All&var-table=All&from=1726041953341&to=1726086965390

[MrCroxx]

[MrCroxx]

[MrCroxx]

[MrCroxx]

Seems the memory buffer overhead in foyer 0.11 caused heavier operator memory eviction in RW.

[Li0k]

Seems the memory buffer overhead in foyer 0.11 caused heavier operator memory eviction in RW.

Can you share the memory consumption metrics from the foyer side ?

[MrCroxx]

Seems the memory buffer overhead in foyer 0.11 caused heavier operator memory eviction in RW.

Can you share the memory consumption metrics from the foyer side ?

Actually, I don't think there is a metrics showing it. It is caused by a temporary buffer that was introduced in 0.11. I'm trying to remove it when there is no compression needs.

[MrCroxx]

Based on the memory prof. The IoBufferPool is the murderer.

collapsed.zip

[MrCroxx]

Solved. The regression is caused by the misleading argument buffer_threshold in v0.11. I'll rename it.