[
https://issues.apache.org/jira/browse/TS-2137?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Yunkai Zhang updated TS-2137:
-----------------------------
Description:
pthread_cond_signal/wait is used in several places in ATS, including but not
limited:
1) Logging system.
2)
As we known, pthread_cond_signal() need to take lock, it'll cause more context
switch than eventfd.
The logging speed is very important for us which will affect RT, it is worth
optimizing.
I can confirm that evenfd is supported in Linux platform.
Let's optimize it in Linux firstly.
===
I have wrote a simple benchmark program to compare the speed of eventfd and
pthread_cond notify mechanism.
You can get the benchmark program from attachment:
https://issues.apache.org/jira/secure/attachment/12598570/eventfd_vs_pthread_cond_benchmark.tar.gz,
and test it by yourself if interesting.
What the program do is that:
1) Create 10 producer threads and 1 consumer thread.
2) Producer threads will notify consumer concurrently in a loop.
3) The consumer thread will receive notification up to a given loop times.
Here is my testing result, the command line argument is loop times(1000000) of
consumer:
{code}
[eventfd_vs_pthread_cond_benchmark]$ ls
main.cc Makefile
[eventfd_vs_pthread_cond_benchmark]$ make
g++ -g -o eventfd_test main.cc -DHAVE_EVENTFD -lpthread
g++ -g -o pthread_cond_test main.cc -lpthread
[eventfd_vs_pthread_cond_benchmark]$ ls
eventfd_test main.cc Makefile pthread_cond_test
[eventfd_vs_pthread_cond_benchmark]$ time ./eventfd_test 1000000
real 0m11.644s
user 0m1.517s
sys 1m31.179s
[eventfd_vs_pthread_cond_benchmark]$ time ./pthread_cond_test 1000000
real 0m57.438s
user 0m30.152s
sys 6m7.289s
{code}
We can see that eventfd is about 5 times faster than pthread_cond notify
mechanism.
was:
At the end of lifetime of each http connection, the logging action will call
pthread_cond_signal() to notify flush threads.
As we known, pthread_cond_signal() need to take lock, it'll cause more context
switch than eventfd.
The logging speed is very important for us which will affect RT, it is worth
optimizing.
I can confirm that evenfd is supported in Linux platform.
Let's optimize it in Linux firstly.
===
I have wrote a simple benchmark program to compare the speed of eventfd and
pthread_cond notify mechanism.
You can get the benchmark program from attachment:
https://issues.apache.org/jira/secure/attachment/12598570/eventfd_vs_pthread_cond_benchmark.tar.gz,
and test it by yourself if interesting.
What the program do is that:
1) Create 10 producer threads and 1 consumer thread.
2) Producer threads will notify consumer concurrently in a loop.
3) The consumer thread will receive notification up to a given loop times.
Here is my testing result, the command line argument is loop times(1000000) of
consumer:
{code}
[eventfd_vs_pthread_cond_benchmark]$ ls
main.cc Makefile
[eventfd_vs_pthread_cond_benchmark]$ make
g++ -g -o eventfd_test main.cc -DHAVE_EVENTFD -lpthread
g++ -g -o pthread_cond_test main.cc -lpthread
[eventfd_vs_pthread_cond_benchmark]$ ls
eventfd_test main.cc Makefile pthread_cond_test
[eventfd_vs_pthread_cond_benchmark]$ time ./eventfd_test 1000000
real 0m11.644s
user 0m1.517s
sys 1m31.179s
[eventfd_vs_pthread_cond_benchmark]$ time ./pthread_cond_test 1000000
real 0m57.438s
user 0m30.152s
sys 6m7.289s
{code}
We can see that eventfd is about 5 times faster than pthread_cond notify
mechanism.
> Use eventfd instread of pthread signal/wait in ATS
> --------------------------------------------------
>
> Key: TS-2137
> URL: https://issues.apache.org/jira/browse/TS-2137
> Project: Traffic Server
> Issue Type: Improvement
> Components: Logging
> Reporter: Yunkai Zhang
> Assignee: Yunkai Zhang
> Fix For: 3.5.0
>
> Attachments:
> 0001-TS-2137-Use-eventfd-instread-of-pthread-signal-wait-.patch,
> 0001-TS-2137-Use-eventfd-instread-of-pthread-signal-wait-.V2.patch,
> eventfd_vs_pthread_cond_benchmark.tar.gz
>
>
> pthread_cond_signal/wait is used in several places in ATS, including but not
> limited:
> 1) Logging system.
> 2)
> As we known, pthread_cond_signal() need to take lock, it'll cause more
> context switch than eventfd.
> The logging speed is very important for us which will affect RT, it is worth
> optimizing.
> I can confirm that evenfd is supported in Linux platform.
> Let's optimize it in Linux firstly.
> ===
> I have wrote a simple benchmark program to compare the speed of eventfd and
> pthread_cond notify mechanism.
> You can get the benchmark program from attachment:
> https://issues.apache.org/jira/secure/attachment/12598570/eventfd_vs_pthread_cond_benchmark.tar.gz,
> and test it by yourself if interesting.
> What the program do is that:
> 1) Create 10 producer threads and 1 consumer thread.
> 2) Producer threads will notify consumer concurrently in a loop.
> 3) The consumer thread will receive notification up to a given loop times.
> Here is my testing result, the command line argument is loop times(1000000)
> of consumer:
> {code}
> [eventfd_vs_pthread_cond_benchmark]$ ls
> main.cc Makefile
> [eventfd_vs_pthread_cond_benchmark]$ make
> g++ -g -o eventfd_test main.cc -DHAVE_EVENTFD -lpthread
> g++ -g -o pthread_cond_test main.cc -lpthread
> [eventfd_vs_pthread_cond_benchmark]$ ls
> eventfd_test main.cc Makefile pthread_cond_test
> [eventfd_vs_pthread_cond_benchmark]$ time ./eventfd_test 1000000
> real 0m11.644s
> user 0m1.517s
> sys 1m31.179s
> [eventfd_vs_pthread_cond_benchmark]$ time ./pthread_cond_test 1000000
> real 0m57.438s
> user 0m30.152s
> sys 6m7.289s
> {code}
> We can see that eventfd is about 5 times faster than pthread_cond notify
> mechanism.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
