[jira] [Updated] (IGNITE-4766) Relax worker thread wakeup logic in StipedExecutor

2017-05-26 Thread Vladimir Ozerov (JIRA)

 [ 
https://issues.apache.org/jira/browse/IGNITE-4766?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Vladimir Ozerov updated IGNITE-4766:

Attachment: IGNITE_4766.patch

> Relax worker thread wakeup logic in StipedExecutor
> --
>
> Key: IGNITE-4766
> URL: https://issues.apache.org/jira/browse/IGNITE-4766
> Project: Ignite
>  Issue Type: Task
>  Components: general
>Affects Versions: 1.9
>Reporter: Vladimir Ozerov
>Assignee: Vladimir Ozerov
> Fix For: 2.1
>
> Attachments: IGNITE_4766.patch
>
>
> *Problem*
> Worker threads in {{StripedExecutor}} have {{parked}} state flag. When set to 
> {{true}} NIO threads will call {{LockSupport.unpark}} after submitting new 
> task to a queue.
> The problem is that this flag is only changed by worker thread. Thus, if NIO 
> worker submitted a new task and called {{unpark}}, worker will clean up 
> {{parked}} state only after real wake up what may take microseconds. All 
> subsequent submits from NIO thread occurred during this time will also invoke 
> {{unpark}} thus generating high contention on the underlying OS primitives. 
> Namely, condition variables which ultimately delegates to {{futex}} on Linux.
> *Solution*
> NIO threads must be able to clear {{parked}} state as well. Some CAS magic 
> will be required there. This way we will minimize number of {{unapark}} calls.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)


[jira] [Updated] (IGNITE-4766) Relax worker thread wakeup logic in StipedExecutor

2017-04-06 Thread Vladimir Ozerov (JIRA)

 [ 
https://issues.apache.org/jira/browse/IGNITE-4766?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Vladimir Ozerov updated IGNITE-4766:

Fix Version/s: (was: 2.0)
   2.1

> Relax worker thread wakeup logic in StipedExecutor
> --
>
> Key: IGNITE-4766
> URL: https://issues.apache.org/jira/browse/IGNITE-4766
> Project: Ignite
>  Issue Type: Task
>  Components: general
>Affects Versions: 1.9
>Reporter: Vladimir Ozerov
>Assignee: Vladimir Ozerov
> Fix For: 2.1
>
>
> *Problem*
> Worker threads in {{StripedExecutor}} have {{parked}} state flag. When set to 
> {{true}} NIO threads will call {{LockSupport.unpark}} after submitting new 
> task to a queue.
> The problem is that this flag is only changed by worker thread. Thus, if NIO 
> worker submitted a new task and called {{unpark}}, worker will clean up 
> {{parked}} state only after real wake up what may take microseconds. All 
> subsequent submits from NIO thread occurred during this time will also invoke 
> {{unpark}} thus generating high contention on the underlying OS primitives. 
> Namely, condition variables which ultimately delegates to {{futex}} on Linux.
> *Solution*
> NIO threads must be able to clear {{parked}} state as well. Some CAS magic 
> will be required there. This way we will minimize number of {{unapark}} calls.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)


[jira] [Updated] (IGNITE-4766) Relax worker thread wakeup logic in StipedExecutor

2017-03-02 Thread Vladimir Ozerov (JIRA)

 [ 
https://issues.apache.org/jira/browse/IGNITE-4766?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Vladimir Ozerov updated IGNITE-4766:

Affects Version/s: (was: 1.0)
   1.9

> Relax worker thread wakeup logic in StipedExecutor
> --
>
> Key: IGNITE-4766
> URL: https://issues.apache.org/jira/browse/IGNITE-4766
> Project: Ignite
>  Issue Type: Task
>  Components: general
>Affects Versions: 1.9
>Reporter: Vladimir Ozerov
>Assignee: Vladimir Ozerov
> Fix For: 2.0
>
>
> *Problem*
> Worker threads in {{StripedExecutor}} have {{parked}} state flag. When set to 
> {{true}} NIO threads will call {{LockSupport.unpark}} after submitting new 
> task to a queue.
> The problem is that this flag is only changed by worker thread. Thus, if NIO 
> worker submitted a new task and called {{unpark}}, worker will clean up 
> {{parked}} state only after real wake up what may take microseconds. All 
> subsequent submits from NIO thread occurred during this time will also invoke 
> {{unpark}} thus generating high contention on the underlying OS primitives. 
> Namely, condition variables which ultimately delegates to {{futex}} on Linux.
> *Solution*
> NIO threads must be able to clear {{parked}} state as well. Some CAS magic 
> will be required there. This way we will minimize number of {{unapark}} calls.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)