[
https://issues.apache.org/jira/browse/IGNITE-16879?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17944665#comment-17944665
]
Ivan Bessonov edited comment on IGNITE-16879 at 4/15/25 9:50 AM:
-----------------------------------------------------------------
Tested using the following scenario:
* {{putAll}} transactions with 1000 entries.
* 3 replicas, 1 client with many threads.
* Just enough load to always have checkpoints with _"too many dirty pages"_
reason.
* Throughput with current main - *202254* inserts per second.
* Throughput with the fix - *222561* inserts per second, about *10%* more.
* Graph looks a bit more smooth, but still with large dips that will be
addressed separately.
h3. Graphs
Before:
!Screenshot from 2025-04-15 12-45-07.png|width=582,height=169!
After:
!Screenshot from 2025-04-15 12-45-09.png|width=580,height=168!
was (Author: ibessonov):
Tested using the following scenario:
* {{putAll}} transactions with 1000 entries.
* 3 replicas, 1 client with many threads.
* Just enough load to always have checkpoints with _"too many dirty pages"_
reason.
* Throughput with current main - *202254* inserts per second.
* Throughput with the fix - *222561* inserts per second, about *10%* more.
* Graph looks a bit more smooth, but still with large dips that will be
addressed separately.
h3. Graphs
Before:
!Screenshot from 2025-04-15 12-45-07.png!
After:
!Screenshot from 2025-04-15 12-45-09.png!
> Estimate time-to-write-cp-page and time-to-fsync-cp-page empirically
> --------------------------------------------------------------------
>
> Key: IGNITE-16879
> URL: https://issues.apache.org/jira/browse/IGNITE-16879
> Project: Ignite
> Issue Type: Improvement
> Components: persistence
> Reporter: Roman Puchkovskiy
> Assignee: Ivan Bessonov
> Priority: Major
> Attachments: Screenshot from 2025-04-15 12-45-07.png, Screenshot from
> 2025-04-15 12-45-09.png
>
> Time Spent: 20m
> Remaining Estimate: 0h
>
> The speed-based throttling (implemented in
> SpeedBasedMemoryConsumptionThrottlingStrategy) tries to estimate how much
> time is needed to fully write a checkpoint to disk. To write a page to disk,
> we not only need to pass it to the OS for writing (which takes time), but
> also fsync it to disk (which also takes some time).
> When estimating the time needed to write a checkpoint, the speed-based
> throttle needs an estimation of how much time it is needed to write a page
> and how much time is needed to fsync it. Currently, a naive approximation is
> used: page writing time is equal to page fsyncing time.
> In reality, this is not true, so the speed estimations are off: the more the
> ratio between writing and syncing time differs from 1, the worse our
> estimation is.
> The idea is to actually measure the average time to write a page and fsync a
> page to disk and then use these measured values to obtain better estimations.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
