[
https://issues.apache.org/jira/browse/IGNITE-21379?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17817309#comment-17817309
]
Vladislav Pyatkov edited comment on IGNITE-21379 at 2/14/24 10:22 AM:
----------------------------------------------------------------------
I compare two general approaches: based on IgniteSpinReadWriteLock and based on
ReentrantReadWriteLock. The comparison runs in the dedicated test:
|RW lock test|406|
|Busy lock test|296|
|RW try lock test|358|
|Counter test|200|
This is a latency of 10_000_000 operations in millis. One operation contains
two read lock and two read unlock operations.
There are the same results, but in multi-threaded environments:
|RW lock test|2 135|
|Busy lock test|2 377|
|RW try lock test|2 088|
|Counter test|1 476|
One operation acquires a lock in one thread, then takes a lock and releases it
in another one, finally returning to the original thread and releasing the
first read lock.
Also, I implement IgniteSpinBusyLock based on RW. I ran a load, and this
implementation does not show a significant performance impact.
was (Author: v.pyatkov):
I compare two general approaches: based on IgniteSpinReadWriteLock and based on
ReentrantReadWriteLock. The comparison runs in the dedicated test:
|RW lock test|406|
|Busy lock test|296|
|RW try lock test|358|
|Counter test|200|
This is a latency of 1000 operations in millis. One operation contains two read
lock and two read unlock operations.
There are the same results, but in multi-threaded environments:
|RW lock test|2 135|
|Busy lock test|2 377|
|RW try lock test|2 088|
|Counter test|1 476|
One operation acquires a lock in one thread, then takes a lock and releases it
in another one, finally returning to the original thread and releasing the
first read lock.
Also, I implement IgniteSpinBusyLock based on RW. I ran a load, and this
implementation does not show a significant performance impact.
> Investigate whether currently used busyLocks implementation is fast enough
> --------------------------------------------------------------------------
>
> Key: IGNITE-21379
> URL: https://issues.apache.org/jira/browse/IGNITE-21379
> Project: Ignite
> Issue Type: Improvement
> Reporter: Alexander Lapin
> Assignee: Vladislav Pyatkov
> Priority: Major
> Labels: ignite-3, performance
>
> h3. Motivation
> Seems that our busyLocks (IgniteSpinBusyLock) aren't good enough from the
> performance perspective. Let's compare current implementation with common RW
> locks, CheckpointReadWriteLock, etc. Depending on the results it'll be
> required either to use faster implementation or re-consider busyLock idea
> itself because currently it brings significant performance drop. Given ticket
> is only about initial step - busyLock performance investigation.
> h3. Definition of Done
> * Prepare JMH benchmarks for busyLocks performance investigation.
> * Compare IgniteSpinBusyLock, common RW lock, CheckpointReadWriteLock, etc
> in order to understand whether IgniteSpinBusyLock is fast enough.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
