[
https://issues.apache.org/jira/browse/IGNITE-21234?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ivan Bessonov updated IGNITE-21234:
-----------------------------------
Description:
In a situation where we have "too many dirty pages" we trigger checkpoint and
wait until it starts. This can take seconds, because we have to flush
free-lists before acquiring checkpoint write lock. This can cause severe dips
in performance for no good reason.
I suggest introducing two modes for triggering checkpoints when we have too
many dirty pages: soft threshold and hard threshold.
* soft - trigger checkpoint, but don't wait for its start. Just continue all
operations as usual. Make it like a current threshold - 75% of any existing
memory segment must be dirty.
* hard - trigger checkpoint and wait until it starts. The way it behaves right
now. Make it higher than current threshold - 90% of any existing memory segment
must be dirty.
Maybe we should use different values for thresholds, that should be discussed
during the review
was:
In a situation where we have "too many dirty pages" we trigger checkpoint and
wait until it starts. This can take seconds, because we have to flush
free-lists before acquiring checkpoint write lock. This can cause severe dips
in performance for no good reason.
I suggest introducing two modes for triggering checkpoints when we have too
many dirty pages: soft threshold and hard threshold.
* soft - trigger checkpoint, but don't wait for its start. Just continue all
operations as usual. Make it like a current threshold - 75% of any existing
memory segment must be dirty.
* hard - trigger checkpoint and wait until it starts. The way it behaves right
now. Make it higher than current threshold - 90% of any existing memory segment
must be dirty.
> Checkpoint read lock waits for checkpoint write unlock sometimes
> ----------------------------------------------------------------
>
> Key: IGNITE-21234
> URL: https://issues.apache.org/jira/browse/IGNITE-21234
> Project: Ignite
> Issue Type: Improvement
> Reporter: Ivan Bessonov
> Assignee: Ivan Bessonov
> Priority: Major
> Labels: ignite-3
> Time Spent: 10m
> Remaining Estimate: 0h
>
> In a situation where we have "too many dirty pages" we trigger checkpoint and
> wait until it starts. This can take seconds, because we have to flush
> free-lists before acquiring checkpoint write lock. This can cause severe dips
> in performance for no good reason.
> I suggest introducing two modes for triggering checkpoints when we have too
> many dirty pages: soft threshold and hard threshold.
> * soft - trigger checkpoint, but don't wait for its start. Just continue all
> operations as usual. Make it like a current threshold - 75% of any existing
> memory segment must be dirty.
> * hard - trigger checkpoint and wait until it starts. The way it behaves
> right now. Make it higher than current threshold - 90% of any existing memory
> segment must be dirty.
> Maybe we should use different values for thresholds, that should be discussed
> during the review
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
