[infinispan-issues] [JBoss JIRA] (ISPN-5962) Implementation of RpcManagerImpl.invokeRemotely causes high CPU usage

[Dan Berindei (JIRA)]

     [ https://issues.jboss.org/browse/ISPN-5962?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Dan Berindei updated ISPN-5962:
-------------------------------
    Attachment: CompletableFutureBenchmarks.java


I wrote a JMH benchmark to measure the difference between {{CompletableFuture.get()}}, {{CompletableFuture.get(timeout, unit)}}, and {{CountDownLatch.await()}}. The test appears to show a significant difference between all 3 when the CPU is saturated, with {{CompletableFuture.get()}} being the slowest. 

{noformat}
Benchmark                                                     Mode  Cnt       Score      Error  Units
CompletableFutureBenchmarks.completableFutureGetBusy           thrpt   20  157501.654 ± 8227.728  ops/s
CompletableFutureBenchmarks.completableFutureGetSleeping       thrpt   20    6337.177 ±   37.172  ops/s
CompletableFutureBenchmarks.completableFutureTimedGetBusy      thrpt   20  178088.683 ± 7661.149  ops/s
CompletableFutureBenchmarks.completableFutureTimedGetSleeping  thrpt   20    6331.961 ±   29.550  ops/s
CompletableFutureBenchmarks.countDownLatchAwaitBusy          thrpt   20  194923.024 ± 7048.457  ops/s
CompletableFutureBenchmarks.countDownLatchAwaitSleeping      thrpt   20    6326.857 ±   28.239  ops/s
{noformat}

However, the difference between the two {{CompletableFuture}} methods disappears completely in some runs...

{noformat}
Benchmark                                                     Mode  Cnt       Score      Error  Units
CompletableFutureBenchmarks.completableFutureGetBusy           thrpt   20  173977.615 ± 7435.750  ops/s
CompletableFutureBenchmarks.completableFutureGetSleeping       thrpt   20    6340.609 ±   21.529  ops/s
CompletableFutureBenchmarks.completableFutureTimedGetBusy      thrpt   20  172650.973 ± 6382.962  ops/s
CompletableFutureBenchmarks.completableFutureTimedGetSleeping  thrpt   20    6314.509 ±   30.036  ops/s
{noformat}


> Implementation of RpcManagerImpl.invokeRemotely causes high CPU usage
> ---------------------------------------------------------------------
>
>                 Key: ISPN-5962
>                 URL: https://issues.jboss.org/browse/ISPN-5962
>             Project: Infinispan
>          Issue Type: Enhancement
>          Components: Core
>    Affects Versions: 8.1.0.Beta1
>            Reporter: Sanne Grinovero
>            Assignee: Pedro Ruivo
>         Attachments: CompletableFutureBenchmarks.java
>
>
> The method {{org.infinispan.remoting.rpc.RpcManagerImpl.invokeRemotely(Collection<Address>, ReplicableCommand, RpcOptions)}} is implemented by blocking on a {{CompletableFuture}}, but this then stalls on {{java.util.concurrent.CompletableFuture.waitingGet(boolean)}} by spending a significant amount of CPU time by spinning.
> When implementing RPC calls and having to wait for remote operations, spinning is probably not a good idea. Could we try implementing this in some way to hint towards a more pessimistic lock?



--
This message was sent by Atlassian JIRA
(v6.4.11#64026)