[#LU-8] Use non-blocking connect in ksocklnd

[LU-8] Use non-blocking connect in ksocklnd Created: 22/Oct/10 Updated: 03/Jun/11 Resolved: 03/Jun/11
Status:	Resolved
Project:	Lustre
Component/s:	None
Affects Version/s:	None
Fix Version/s:	Lustre 2.1.0, Lustre 1.8.6

Type:

Bug

Priority:

Major

Reporter:

Robert Read (Inactive)

Assignee:

Liang Zhen (Inactive)

Resolution:

Fixed

Votes:

Labels:

None

Bugzilla ID:

14,846

Rank (Obsolete):

5006

Description

Related to bz 14846.

Comments

Comment by Liang Zhen (Inactive) [ 24/Oct/10 ]

Robert,
which branch should it base on? 14846 is created for 1.6, but I guess we want it on newer version right?

Liang

Comment by Dan Ferber (Inactive) [ 26/Oct/10 ]

They are running 1.8.3 and soon moving to 1.8.4, so you are right - they would like it based on the newer release. Dan.

Comment by Christopher Morrone [ 26/Oct/10 ]

Are you really planning on making the connect non-blocking, or just implement dynamic creation of ksocknal_connd() threads as Andreas suggests?

Comment by Liang Zhen (Inactive) [ 26/Oct/10 ]

I think suggestion from Andreas is quite reasonable, we can make connd threads-pool increase & shrink automatically (just like ptlrpc service threads)

if there is too many active connecting requests, create more threads
if number of connecting requests is under a low water, we can kill some connds

Christopher, if you think it's acceptable, then I can start very soon.

Regards
Liang

Comment by Liang Zhen (Inactive) [ 28/Oct/10 ]

First version of patch

http://review.whamcloud.com/#change,119

Liang

Comment by Liang Zhen (Inactive) [ 02/Nov/10 ]

I've made some changes due to inspection, second version of patch is here: http://review.whamcloud.com/#change,119

Liang

Comment by Robert Read (Inactive) [ 03/Nov/10 ]

Chris, the patch has been inspected and Liang has tested it as much as he can, so it seems we're ready to pass it on to you guys (and also to Oracle). You can pull the 1.8-based
patch directly from Gerrit (the review link above has instructions for several methods of downloading the patch), or, if there are merge issues on your end, we can add a mirror of your chaos tree to review.whamcloud.com and merge the patch onto chaos here. (I do intend to mirror your tree eventually, but things are bit busy right now with SC coming up.)

Please let me or Dan know if that works for you or if you would like to manage this differently.

Comment by Christopher Morrone [ 04/Nov/10 ]

Robert, I am finding the gerrit page a bit confusing. It appears from the comments at the bottom that I only want patch set 2, but there appears to be no way from the information at the top to discern that fact.

It gives every appearance that both patch sets are approved as a whole.

Comment by Robert Read (Inactive) [ 05/Nov/10 ]

I spoke with Chris yesterday to sort out the issues with Gerrit. He's going to pull the patch and try this on his tree next.

Comment by Dan Ferber (Inactive) [ 09/Nov/10 ]

Liang, can you put a short comment in bz 14846 just to say you are working on a related fix and give the Jira pointer to this bug. Thanks.

Comment by Dan Ferber (Inactive) [ 12/Nov/10 ]

Status: Next step is for LLNL to test in their environment.

Comment by Christopher Morrone [ 06/Dec/10 ]

I added a comment on the patch in gerrit.

Comment by Liang Zhen (Inactive) [ 07/Dec/10 ]

Chris, I've updated the patch, briefly I just removed the interval for creating new thread. I didn't reserve more threads for creating new threads, because we have already reserved one thread for accepting, which is unlikely to be blocked for long time (the worst case is 10 seconds, which is unlikely to happen). But I can change it to have a dedicated thread for creating new thread if you still think it's necessary.

Liang

Comment by Liang Zhen (Inactive) [ 27/Jan/11 ]

Chris, have you got any chance to try this patch? I'm thinking about push it into upstream if possible.

Thanks
Liang

Comment by Christopher Morrone [ 27/Jan/11 ]

I am sorry, I have not had the time to test this.

Comment by Liang Zhen (Inactive) [ 03/Feb/11 ]

Patch landed on 1.8.*
I will port it to 2.* and push it to upstream

Comment by Build Master (Inactive) [ 01/Apr/11 ]

Integrated in reviews-centos5 #633
~~LU-8~~ dynamically grow/shrink socklnd connd threads