[JBoss JIRA] (ISPN-6003) Reduce number of allocations
by RH Bugzilla Integration (JIRA)
[ https://issues.jboss.org/browse/ISPN-6003?page=com.atlassian.jira.plugin.... ]
RH Bugzilla Integration commented on ISPN-6003:
-----------------------------------------------
Sebastian Łaskawiec <slaskawi(a)redhat.com> changed the Status of [bug 1272390|https://bugzilla.redhat.com/show_bug.cgi?id=1272390] from POST to MODIFIED
> Reduce number of allocations
> ----------------------------
>
> Key: ISPN-6003
> URL: https://issues.jboss.org/browse/ISPN-6003
> Project: Infinispan
> Issue Type: Task
> Components: Core, Server
> Reporter: Dan Berindei
> Assignee: Dan Berindei
> Labels: performance
> Fix For: 8.1.0.Final
>
>
> Profiling revealed some allocations that are easy to remove:
> * The HotRod operations factory stores a list of flags in a thread-local. The thread-local can be removed, and the flags can be stored in an {{int}}.
> * JGroupsTransport copies the list of members to check if a broadcast should be sent as a unicast, and the copy is then discarded.
> * ExtendedByteBuf could use {{Array.empty}} instead of {{Array[Byte]()}}.
> * DecoratedCache could avoid calling {{Arrays.asList()}}.
--
This message was sent by Atlassian JIRA
(v6.4.11#64026)
8 years, 5 months
[JBoss JIRA] (ISPN-5857) HotRod client should use the consistent hash in replicated mode
by RH Bugzilla Integration (JIRA)
[ https://issues.jboss.org/browse/ISPN-5857?page=com.atlassian.jira.plugin.... ]
RH Bugzilla Integration commented on ISPN-5857:
-----------------------------------------------
Sebastian Łaskawiec <slaskawi(a)redhat.com> changed the Status of [bug 1272390|https://bugzilla.redhat.com/show_bug.cgi?id=1272390] from POST to MODIFIED
> HotRod client should use the consistent hash in replicated mode
> ---------------------------------------------------------------
>
> Key: ISPN-5857
> URL: https://issues.jboss.org/browse/ISPN-5857
> Project: Infinispan
> Issue Type: Bug
> Components: Server
> Affects Versions: 8.0.1.Final, 8.1.0.Alpha1
> Reporter: Dan Berindei
> Assignee: Dan Berindei
> Priority: Critical
> Labels: hotrod
> Fix For: 8.1.0.Alpha2, 8.1.0.Final
>
>
> The HotRod server assumes that in replicated mode the server accessed by the client doesn't matter, so it doesn't send a consistent hash in the topology updates.
> However, since replicated mode is now implemented on top of distributed mode, hitting the primary owner or another node makes a big difference. We should change the server to send the consistent hash for replicated caches as well.
--
This message was sent by Atlassian JIRA
(v6.4.11#64026)
8 years, 5 months
[JBoss JIRA] (ISPN-5883) Node can apply new topology after sending status response
by Dan Berindei (JIRA)
[ https://issues.jboss.org/browse/ISPN-5883?page=com.atlassian.jira.plugin.... ]
Dan Berindei commented on ISPN-5883:
------------------------------------
I missed one case in my pull request: the node applying a topology update after sending a status response could be the coordinator itself. If the coordinator after the merge is also the coordinator of a pre-merge partition, waiting for the view and checking the coordinator doesn't help.
We could fix this either on the ClusterTopologyManager side or on the LocalTopologyManager side:
* Cluster status recovery could include the topologies sent out by the coordinator, but not yet received by the members (from ClusterCacheStatus).
* A member could refuse to apply topology updates with a view id smaller than the view id of the latest cluster status request.
> Node can apply new topology after sending status response
> ---------------------------------------------------------
>
> Key: ISPN-5883
> URL: https://issues.jboss.org/browse/ISPN-5883
> Project: Infinispan
> Issue Type: Bug
> Components: Core, Test Suite - Core
> Affects Versions: 8.0.1.Final, 7.2.5.Final, 8.1.0.Alpha2
> Reporter: Dan Berindei
> Assignee: Dan Berindei
> Priority: Critical
> Labels: testsuite_stability
> Fix For: 8.1.0.Final
>
>
> {{LocalTopologyManagerImpl}} is responsible for sending the {{ClusterTopologyControlCommand(GET_STATUS)}} response, and when it sends the response it doesn't check the current view id against the new coordinator's view id. If the old coordinator already sent a topology update before the merge, that topology update might be processed after sending the status response. The new coordinator will send a topology update with a topology id of {{max(status response topology ids) + 1}}. The node will then process the topology update from the old coordinator, but it will ignore the topology update from the new coordinator with the same topology id.
> This is extra common in the partition handling tests, e.g. {{BasePessimisticTxPartitionAndMergeTest}} subclasses, because the test "injects" the JGroups view on each node serially, and often the 4th node sends the status response before it gets the new view.
> {noformat}
> 22:16:37,776 DEBUG (remote-thread-NodeD-p26-t6:[]) [LocalTopologyManagerImpl] Sending cluster status response for view 10
> // Topology from NodeC
> 22:16:37,778 DEBUG (transport-thread-NodeD-p28-t2:[]) [LocalTopologyManagerImpl] Updating local topology for cache pes-cache: CacheTopology{id=8, rebalanceId=3, currentCH=DefaultConsistentHash{ns=60, owners = (4)[NodeA-37631: 15+15, NodeB-47846: 15+15, NodeC-46467: 15+15, NodeD-30486: 15+15]}, pendingCH=null, unionCH=null, actualMembers=[NodeC-46467, NodeD-30486]}
> // Later, topology from NodeA
> 22:16:37,827 DEBUG (transport-thread-NodeD-p28-t1:[]) [LocalTopologyManagerImpl] Ignoring late consistent hash update for cache pes-cache, current topology is 8: CacheTopology{id=8, rebalanceId=3, currentCH=DefaultConsistentHash{ns=60, owners = (4)[NodeA-37631: 15+15, NodeB-47846: 15+15, NodeC-46467: 15+15, NodeD-30486: 15+15]}, pendingCH=null, unionCH=null, actualMembers=[NodeA-37631, NodeB-47846, NodeC-46467, NodeD-30486]}
> {noformat}
> As a solution, we can delay sending the status response until we have the same view as the coordinator (or a later one). We already check that the sender is the current coordinator before applying a topology update, so this will guarantee that the we don't apply other topology updates from the old coordinator. Since the status request is only sent after the new view was installed, this will not introduce any delays in the vast majority of cases.
--
This message was sent by Atlassian JIRA
(v6.4.11#64026)
8 years, 5 months