Dan Berindei created ISPN-4585: ---------------------------------- Summary: Prioritize commands in the remote executor Key: ISPN-4585 URL: https://issues.jboss.org/browse/ISPN-4585 Project: Infinispan Issue Type: Enhancement Security Level: Public (Everyone can see) Components: Core Affects Versions: 7.0.0.Alpha5 Reporter: Dan Berindei Assignee: Dan Berindei Fix For: 7.0.0.Final The remote executor currently has an unlimited queue of blocked task, but the underlying executor cannot use a queue. With a queue, we wouldn't need to overflow remote commands to the OOB threads, and the OOB threads would be free to process response messages. The problem is that {{ThreadPoolExecutor}} executes tasks in the order they are in the queue. If a node has a remote executor thread pool of 100 threads and receives a prepare(tx1, put(k, v1) comand, then 1000 prepare(tx_i, put(k, v_i)) commands, and finally a commit(tx1) command, the commit(tx1) command will block until all but 99 of the the prepare(tx_i, put(k, v_i)) commands have timed out. I think we could help this by using a {{PriorityBlockingQueue}} for the underlying executor, with commands ordered so that state transfer commands < commit/tx completion notification < prepare/lock. The commit command would still have to wait for one of the prepare commands currently running to time out, but it wouldn't have to wait for all of them. The current code, without a queue, would fill the remote executor and OOB thread pools, and it would discard the commit message (along with most of the prepare commands). The time it would take to process the commit successfully would depend on the timing of the retransmitted messages. Another possible improvement would be to keep track of the commands currently being executed, and always keep some threads free for commands with higher priority. But I'm not sure how easy it would be to do that on top of an injected {{ExecutorService}}. I believe there is also a problem with {{BlockingTaskAwareExecutorServiceImpl.checkForReadyTasks()}} after a topology change. Commands with the new topology id are all unblocked by submitting them to the underlying executor in FIFO order, on a single thread, so {{CallerRunsPolicy}} is not a valid rejection policy here. -- This message was sent by Atlassian JIRA (v6.2.6#6264)