[
https://issues.apache.org/jira/browse/MAPREDUCE-7329?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
chaoli updated MAPREDUCE-7329:
------------------------------
Description:
Recently, we upgrade linux kernel version from 3.x to 4.x. And we find hadoop
pipe task exit with connect timeout which is implemented by PingThread in
HadoopPipes.cc.
!image-2021-03-15-14-37-32-184.png!
After a deep research, we finally find that current ping server won't accept
ping client created socket, which has hidden danger:
# it will cause tcp accept queue full(*default 50*)
# when client close socket, server socket won't call close method, which will
leave too many CLOSE_WAIT socket fd existed(*default 2h*), and accept queue
never cleared.
# Even worse, in 4.x linux kernel version, it will cause tcp drop packet
directly which makes ping client connect time out. While In 3.x linux kernel
version, when accept queue full, client can also make half connection till sync
queue full (*default 2048*), so from client side, ping will aslo work till sync
queue full. And after 3 hours, task will also exit with connect timeout
exception.
To fix this problem, we introduced a *PingSocketCleaner* thread, which will
continuously accept ping socket connect from ping client. When socket close
from client, cleaner thread will detected by inputStream read, then will
finally close socket from sever side.
Refrenced by linux kernel patch:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=5ea8ea2cb7
was:
Recently, we upgrade linux kernel version from 3.x to 4.x. And we find hadoop
pipe task exit with connect timeout which is implemented by PingThread in
HadoopPipes.cc.
!image-2021-03-15-14-37-32-184.png!
After a deep research, we finally find that current ping server won't accept
ping client created socket, which has hidden danger:
# it will cause tcp accept queue full(*default 50*)
# when client close socket, server socket won't call close method, which will
leave too many CLOSE_WAIT socket fd existed(*default 2h*), and accept queue
never cleared.
# Even worse, in 4.x linux kernel version, it will cause tcp drop packet
directly which makes ping client connect time out. While In 3.x linux kernel
version, when accept queue full, client can also make half connection till sync
queue full (*default 2048*), so from client side, ping will aslo work till sync
queue full. And after 3 hours, task will also exit with connect timeout
exception.
To fix this problem, we introduced a *PingSocketCleaner* thread, which will
continuously accept ping socket connect from ping client. When socket close
from client, cleaner thread will detected by inputStream read, then will
finally close socket from sever side.
> HadoopPipes task may fail when linux kernel version change from 3.x to 4.x
> --------------------------------------------------------------------------
>
> Key: MAPREDUCE-7329
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-7329
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Affects Versions: 2.6.0
> Reporter: chaoli
> Priority: Major
> Labels: patch, pull-request-available
> Attachments:
> 0001-MAPREDUCE-7329-HadoopPipes-task-may-fail-when-linux-.patch,
> image-2021-03-15-14-29-49-475.png, image-2021-03-15-14-37-32-184.png
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> Recently, we upgrade linux kernel version from 3.x to 4.x. And we find hadoop
> pipe task exit with connect timeout which is implemented by PingThread in
> HadoopPipes.cc.
> !image-2021-03-15-14-37-32-184.png!
> After a deep research, we finally find that current ping server won't accept
> ping client created socket, which has hidden danger:
> # it will cause tcp accept queue full(*default 50*)
> # when client close socket, server socket won't call close method, which
> will leave too many CLOSE_WAIT socket fd existed(*default 2h*), and accept
> queue never cleared.
> # Even worse, in 4.x linux kernel version, it will cause tcp drop packet
> directly which makes ping client connect time out. While In 3.x linux kernel
> version, when accept queue full, client can also make half connection till
> sync queue full (*default 2048*), so from client side, ping will aslo work
> till sync queue full. And after 3 hours, task will also exit with connect
> timeout exception.
> To fix this problem, we introduced a *PingSocketCleaner* thread, which will
> continuously accept ping socket connect from ping client. When socket close
> from client, cleaner thread will detected by inputStream read, then will
> finally close socket from sever side.
> Refrenced by linux kernel patch:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=5ea8ea2cb7
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: mapreduce-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: mapreduce-issues-h...@hadoop.apache.org
