[Ocfs2-users] Other problems with OCFS2 hanging on writes

Mon Nov 19 18:07:38 PST 2012

Hello,
A month ago, following the advice from Srini, we modified our localalloc to be 2048.   It made sense at that time because our disk fragmentation was showing many "2048 clusters" of contiguous space available.
# ./stat_sysdir-analyze.sh stat_sysdir-output.txt-201210260843 Number |of     |clust. | Contiguous cluster size--------------------------------   3981 510 and smaller   4929 511   2393 1024  63400 2048   2682 4096   2415 8192    430 16384 and bigger
It all worked very well and saved the day!

Today, we began to have more issues with our disk writes hanging.  This time though it might be a bit different ...
First of all, while the filesystem is even more fragmented than before, there are still many "2048 clusters" available:
# ./stat_sysdir-analyze.sh stat_sysdir-output.txt-201211191829 Number |of     |clust. | Contiguous cluster size--------------------------------   4407 510 and smaller  50270 511  13609 1024  10338 2048    649 4096    347 8192    641 16384 and bigger
and then, strangely, the problem appeared only every 20 minutes, lasting 2-3 minutes each time.
Could the problem be related with some internal process in OCFS2 running every 20 minutes?   Orphan scans?
Or could the problem be due to the excessive fragmentation?   Should I change the localalloc size to be lower than 511 clusters instead of 2048 (i.e.  localalloc=3)?   Would that actually worsen the fragmentation?     For information, I don't think the filesystem is over-utilized.  There are only 12TB used on a 18TB filesystem so that should leave enough space to properly deal with fragmentation.

.... and finally, the problem suddenly disappeared after I shutdown 2 of the 3 nodes of the cluster.  I re-enabled the 2 nodes 30 minutes later and the problem didn't come back.  On the other hand, most of the employees using that filesystem finished their work day so the load is quite lower now.

Any idea of what happened?
Thanks in advance,
Jeff

p.s.  I attached the small script I wrote to analyze the data gathered from Srini's script (https://oss.oracle.com/~seeda/misc/stat_sysdir.sh)

From: SKempinski at sjrwmd.com
To: jpaterson23 at hotmail.com; ocfs2-users at oss.oracle.com
Subject: RE: [Ocfs2-users] OCFS2 hanging on writes -- SOLVED
Date: Wed, 7 Nov 2012 23:44:06 +0000

Thank you for providing this write-up, Jeff.  I think you explained it very well.

This could be the issue we are experiencing, so I've implemented the localalloc mount option.  Additionally, we have begun to move some of the data off to a new filesystem.  For the past few days we have not seen the slowdown, so, for now at least, we have
 a solution.

Thank you, again.

Scott

From: ocfs2-users-bounces at oss.oracle.com [ocfs2-users-bounces at oss.oracle.com] on behalf of Jeff Paterson [jpaterson23 at hotmail.com]

Sent: Wednesday, October 31, 2012 9:24 PM

To: Scott Kempinski; ocfs2-users at oss.oracle.com

Subject: Re: [Ocfs2-users] OCFS2 hanging on writes -- SOLVED

Hello Scott,

I had help from an Oracle developer, Srinivas, and he fixed my issue.   Thanks again Srini !!

Disclaimer: I am not technically knowledgable of the OCFS2 filesystem so I will explain what I understood from a discussion with Srini.

The main issue was that my filesystem is getting more and more fragmented and the default write pre-allocation window (localalloc bitmap) was
 set too big for such fragmented filesystem.  For what I understood, when you are doing writes on a OCFS2 filesystem, the filesystem reserves a chunk of space before beginning to write onto the filesystem.  Even if you are
 writing a very small file, the filesystem will always reserve that chunk of space (I assume this helps reduce fragmentation ?!)

According to my filesystem setup, the size of the pre-allocated chunks was set at 136 MB so it meant the filesystem needed to find 136 MB of contiguous space every time a write was being done.  That caused delays because it
 had hard time finding them ...

Srini showed me how to reduce the pre-allocated chunks size (localalloc bitmap) to a smaller size (16 MB instead of 136 MB) and, since then, everything works as new.   The solution to my problem was to add localalloc=16 to
 my filesystem mount options, umount/mount the filesystem and everything was fixed.

[root at fileserv01 ~]# grep tier2-ocfs2 /etc/fstab 

LABEL=tier2-ocfs2 /tier2-ocfs2 ocfs2 _netdev,nodev,noatime,errors=panic,data=writeback,noacl,nouser_xattr,commit=60,localalloc=16 0
 0

For info, you can view your current localalloc setting by looking at the fs_state in the debugfs.

You first need to mount the virtual debugfs filesystem if it's not already mounted: 

[root at fileserv01 ~]# grep debugfs /etc/fstab 

debugfs /sys/kernel/debug debugfs 0 0

My localalloc settings before the change:

[root at fileserv01 ~]# grep "LocalAlloc =" /sys/kernel/debug/ocfs2/*/fs_state
LocalAlloc => State: 1  Descriptor: 0  Size:
17441 bits  Default: 29696 bits

My localalloc settings after the change

[root at fileserv01 ~]# grep "LocalAlloc =" /sys/kernel/debug/ocfs2/*/fs_state
LocalAlloc => State: 1  Descriptor: 0  Size:
2048 bits  Default: 2048 bits

What you will be missing from my above post is the analysis from Srini where he found that they were not many "136 MB" chunks of contiguous space on my filesystem and therefore that tuning was definitely going to help.

I hope this post may help you and others.

Jeff

p.s.   sadly, there is currently no defragmentation tool for OCFS2

From: SKempinski at sjrwmd.com

To: jpaterson23 at hotmail.com; ocfs2-users at oss.oracle.com

Subject: RE: [Ocfs2-users] OCFS2 hanging on writes

Date: Wed, 31 Oct 2012 12:30:00 +0000

Jeff,

Have you found a resolution to this issue?

Lately we've been experiencing intermittent freezing, so I'm curious to hear more about your issue.

Thanks, Scott

From: ocfs2-users-bounces at oss.oracle.com [ocfs2-users-bounces at oss.oracle.com] on behalf of Jeff Paterson [jpaterson23 at hotmail.com]

Sent: Thursday, October 25, 2012 9:32 PM

To: ocfs2-users at oss.oracle.com

Subject: [Ocfs2-users] OCFS2 hanging on writes

Hello,

I would need help with our OCFS2 (1.8.0) filesystem.  We are having problems with it since a couple days.  When we write onto it, it hangs.

The "hanging pattern" is easily reproductible.  If I write a 1GB file on the filesystem, it does the following:
        - write ~200 MB of data on the disk in 1 second
        - freeze for about 10 seconds
        - write ~200 MB of data on the disk in 1 second
        - freeze for about 10 seconds
        - write ~200 MB of data on the disk in 1 second
        - freeze for about 10 seconds
        (and so on)

When the freezes occur:
        - other writes operations (from other processes) on the same node also freeze
        - writes operations on other nodes are not affected by the freezes on another node

Read operations (on any cluster node, even the one with frozen writes) don't seem to be affected by the freezes.  One sure thing, read operations alone don't
 cause the filesystem freeze.

For info, before the problem began to appear we could sustain 640 MB/s writes without any freeze.

I tried to mount the filesystem on a single node to avoid issues that could happen with inter-node communications and the problem was still there.

Filesystem details

The filesystem has 18 TB and it is currently 72% full.Mount options are the following: rw,nodev,_netdev,noatime,errors=panic,data=writeback,noacl,nouser_xattr,commit=60,heartbeat=localAll Features: backup-super strict-journal-super sparse extended-slotmap inline-data metaecc indexed-dirs refcount discontig-bg unwritten

There is nothing special in the systems logs beside application errors caused by the freezes.

Would a fsck.ocfs2 help?   How long would it take for 18 TB?

Is there a flag I can enable in debugfs.ocfs2 to get a better idea of what is happening and why it is freezing like that?

Any help would be greatly appreciated.

Thanks in advance,

Jeff

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://oss.oracle.com/pipermail/ocfs2-users/attachments/20121119/c1f9f6df/attachment-0001.html 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: stat_sysdir-analyze.sh
Type: application/octet-stream
Size: 924 bytes
Desc: not available
Url : http://oss.oracle.com/pipermail/ocfs2-users/attachments/20121119/c1f9f6df/attachment-0001.obj 