[
https://issues.apache.org/jira/browse/IGNITE-21464?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17815130#comment-17815130
]
Roman Puchkovskiy commented on IGNITE-21464:
--------------------------------------------
I made a dirty test to check how much one thread switch might cost in the
Embedded mode. I applied the following patch:
{{a/modules/replicator/src/main/java/org/apache/ignite/internal/replicator/ReplicaManager.java
b/modules/replicator/src/main/java/org/apache/ignite/internal/replicator/ReplicaManager.java}}
{{index ba75082063..8e017c88cd 100644}}
{{---
a/modules/replicator/src/main/java/org/apache/ignite/internal/replicator/ReplicaManager.java}}
{{+++
b/modules/replicator/src/main/java/org/apache/ignite/internal/replicator/ReplicaManager.java}}
{{@@ -246,8 +246,13 @@ public class ReplicaManager extends
AbstractEventProducer<LocalReplicaEvent, Loc}}
{{ }}
{{ ReplicaRequest request = (ReplicaRequest) message;}}
{{ }}
{{- ExecutorService stripeExecutor =
ReplicationGroupStripes.stripeFor(request.groupId(), requestsExecutor);}}
{{- stripeExecutor.execute(() -> handleReplicaRequest(request,
senderConsistentId, correlationId));}}
{{+ if
(Thread.currentThread().getName().contains("MessagingService-inbound")) {}}
{{+ // The request is from the network.}}
{{+ ExecutorService stripeExecutor =
ReplicationGroupStripes.stripeFor(request.groupId(), requestsExecutor);}}
{{+ stripeExecutor.execute(() -> handleReplicaRequest(request,
senderConsistentId, correlationId));}}
{{+ } else {}}
{{+ handleReplicaRequest(request, senderConsistentId,
correlationId);}}
{{+ }}}
{{ }}}
{{ }}
{{ private void handleReplicaRequest(ReplicaRequest request, String
senderConsistentId, @Nullable Long correlationId) {}}
It basically makes sure that, if a ReplicaRequest comes from the same node (and
not from the network), we don't switch it to another thread and just proceed in
the current thread.
For a scenario of 300k PUTs, this caused throughput to increase from 2976 to
3168 op/sec.
For a scenario of 300k GETs, this increased throughput from 15921 to 19293
op/sec.
Both scenarios are about Embedded mode (so no Thin Client) on a 1-node cluster.
> Reduce the number of context switches
> -------------------------------------
>
> Key: IGNITE-21464
> URL: https://issues.apache.org/jira/browse/IGNITE-21464
> Project: Ignite
> Issue Type: Improvement
> Reporter: Vladislav Pyatkov
> Priority: Major
> Labels: ignite-3
>
> h3. Motivation
> We move the flow of execution from one thread to another. Each time, it
> decreases the performance of an operation, especially in the case of embedded
> mode.
> Probably, we can avoid context switching in particular conditions (embed mode
> or any other circumstance where the execution thread is already fit for the
> operation).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
