[jira] [Commented] (TS-3044) linux native AIO should use eventfd if available to signal thread
[ https://issues.apache.org/jira/browse/TS-3044?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14157141#comment-14157141 ] ASF subversion and git services commented on TS-3044: - Commit 86295176ccb5b7ef5c96686b398733e1edf6f2ee in trafficserver's branch refs/heads/master from [~jplevyak] [ https://git-wip-us.apache.org/repos/asf?p=trafficserver.git;h=8629517 ] TS-3044: Use eventfd in AIO_MODE_NATIVE if available linux native AIO should use eventfd if available to signal thread - Key: TS-3044 URL: https://issues.apache.org/jira/browse/TS-3044 Project: Traffic Server Issue Type: Improvement Components: Cache Reporter: John Plevyak Assignee: Phil Sorber Fix For: 5.2.0 Attachments: native-aio-eventfd.patch linux native AIO has the ability to signal the event thread to get off the poll and service the disk via the io_set_eventfd() call. linux native AIO scales better than the thread-based IO, but the current implementation can introduce delays on lightly loaded systems because of the thread is waiting on epoll(). This can be remedied by using io_set_eventfd -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (TS-3044) linux native AIO should use eventfd if available to signal thread
[ https://issues.apache.org/jira/browse/TS-3044?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14154972#comment-14154972 ] Phil Sorber commented on TS-3044: - Ok, you've convinced me. I am +1 on this patch. If you want to commit this as is, I will fix the perror() in the AIO_MODE_NATIVE code. linux native AIO should use eventfd if available to signal thread - Key: TS-3044 URL: https://issues.apache.org/jira/browse/TS-3044 Project: Traffic Server Issue Type: Improvement Components: Cache Reporter: John Plevyak Assignee: Phil Sorber Fix For: 5.2.0 Attachments: native-aio-eventfd.patch linux native AIO has the ability to signal the event thread to get off the poll and service the disk via the io_set_eventfd() call. linux native AIO scales better than the thread-based IO, but the current implementation can introduce delays on lightly loaded systems because of the thread is waiting on epoll(). This can be remedied by using io_set_eventfd -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (TS-3044) linux native AIO should use eventfd if available to signal thread
[
https://issues.apache.org/jira/browse/TS-3044?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14153806#comment-14153806
]
Phil Sorber commented on TS-3044:
-
[~jplevyak]
I really like this approach. This is something I feel we should do with the
threaded AIO as well.
A couple of comments:
Why perror here and not some ATS built in logging?
{noformat}
+ perror(io_getevents);
{noformat}
Does lowering the AIO_PERIOD make sense globally if we have a case where we may
still not use eventfd?
{noformat}
-#define AIO_PERIOD-HRTIME_MSECONDS(4)
+#define AIO_PERIOD-HRTIME_MSECONDS(10)
--
+#ifdef HAVE_EVENTFD
+ io_set_eventfd(op-aiocb, t-evfd);
+#endif
{noformat}
linux native AIO should use eventfd if available to signal thread
-
Key: TS-3044
URL: https://issues.apache.org/jira/browse/TS-3044
Project: Traffic Server
Issue Type: Improvement
Components: Cache
Reporter: John Plevyak
Assignee: Phil Sorber
Fix For: 5.2.0
Attachments: native-aio-eventfd.patch
linux native AIO has the ability to signal the event thread to get off the
poll and service the disk via the io_set_eventfd() call. linux native AIO
scales better than the thread-based IO, but the current implementation can
introduce delays on lightly loaded systems because of the thread is waiting
on epoll(). This can be remedied by using io_set_eventfd
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
[jira] [Commented] (TS-3044) linux native AIO should use eventfd if available to signal thread
[ https://issues.apache.org/jira/browse/TS-3044?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14154210#comment-14154210 ] John Plevyak commented on TS-3044: -- The perror is from the original AIO_MODE_NATIVE code. I was just following the style as this was a minimal patch just to add the eventfd handling. I agree that we should change that to standard ATS errors. For most unix/linux installations waiting for less than 10msec is the same as waiting for 0msecs, and can result in busy spinning. The iocore has a minimum wait time, so HRTIME_MSECONDS(4) is disingenuous as well as being a poor idea (if it actually was obeyed). linux native AIO should use eventfd if available to signal thread - Key: TS-3044 URL: https://issues.apache.org/jira/browse/TS-3044 Project: Traffic Server Issue Type: Improvement Components: Cache Reporter: John Plevyak Assignee: Phil Sorber Fix For: 5.2.0 Attachments: native-aio-eventfd.patch linux native AIO has the ability to signal the event thread to get off the poll and service the disk via the io_set_eventfd() call. linux native AIO scales better than the thread-based IO, but the current implementation can introduce delays on lightly loaded systems because of the thread is waiting on epoll(). This can be remedied by using io_set_eventfd -- This message was sent by Atlassian JIRA (v6.3.4#6332)
[jira] [Commented] (TS-3044) linux native AIO should use eventfd if available to signal thread
[ https://issues.apache.org/jira/browse/TS-3044?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14110181#comment-14110181 ] John Plevyak commented on TS-3044: -- Assigned to weijin for review as he was in charge of the native linux AIO and can assess the impact. As I remember, this isn't enabled by default because of the latency concerns by Leif. With this patch, if the latency concerns are addressed, we might want to enable this feature by default. linux native AIO should use eventfd if available to signal thread - Key: TS-3044 URL: https://issues.apache.org/jira/browse/TS-3044 Project: Traffic Server Issue Type: Improvement Components: Cache Reporter: John Plevyak Assignee: weijin Attachments: native-aio-eventfd.patch linux native AIO has the ability to signal the event thread to get off the poll and service the disk via the io_set_eventfd() call. linux native AIO scales better than the thread-based IO, but the current implementation can introduce delays on lightly loaded systems because of the thread is waiting on epoll(). This can be remedied by using io_set_eventfd -- This message was sent by Atlassian JIRA (v6.2#6252)
[jira] [Commented] (TS-3044) linux native AIO should use eventfd if available to signal thread
[ https://issues.apache.org/jira/browse/TS-3044?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=14110206#comment-14110206 ] weijin commented on TS-3044: Nice work, John. And I can`t wait to test it. :) linux native AIO should use eventfd if available to signal thread - Key: TS-3044 URL: https://issues.apache.org/jira/browse/TS-3044 Project: Traffic Server Issue Type: Improvement Components: Cache Reporter: John Plevyak Assignee: weijin Fix For: 5.2.0 Attachments: native-aio-eventfd.patch linux native AIO has the ability to signal the event thread to get off the poll and service the disk via the io_set_eventfd() call. linux native AIO scales better than the thread-based IO, but the current implementation can introduce delays on lightly loaded systems because of the thread is waiting on epoll(). This can be remedied by using io_set_eventfd -- This message was sent by Atlassian JIRA (v6.2#6252)
