Date:	Mon, 7 Jan 2008 15:58:29 +0100
From:	Guillaume Laurès <guillaume-laures@...f.fr>
To:	linux-kernel@...r.kernel.org
Subject: PROBLEM: I/O scheduler problem with an 8 SATA disks raid 5 under heavy load  ?

Hello,

Dear kernel developers, my dmesg asked me to report this, so here I  
go ;)
Here is what I found in my dmesg: "anticipatory: forced dispatching  
is broken (nr_sorted=1), please report this".

- First, let's talk about the machine: it's quite pushed so maybe the  
cause is me doing something wrong rather than a bug in the kernel.

I got this alert on a dual core amd64 xen host. It has 8 SATA drives  
making a raid 5 array. This array makes a virtual block device for  
one of the virtual machines:  an Openfiler appliance. Openfiler then  
manages logical volumes on this device including an XFS partition  
shared via NFS. 2 MythTV hosts continuously write MPEG2 tv shows on  
it (1 to 4Gb each).
Still following ? Here is a summary: MPEG2 files -> NFS -> XFS -> LVM  
-> Xen VBD -> RAID 5 -> 8x SATA disks.

- Next, the symptoms.

This setup is only 2 weeks old. Behavior was quite good, except for  
some unexplained failures from the sata_nv attached disks. Not always  
from the same disk. Never from any disks attached through the  
sata_sil HBA.
Eventually a second disk would go down before the end of the raid  
reconstruction (still a sata_nv attached one).
Since the disks showed nothing wrong with smartmontools I re-added  
them each time. So far the raid array was strong enough to be fully  
recovered, mdadm --force and xfs_check are my friends ;-)
It seems to happen more often now that the XFS partition is quite  
heavily fragmented, and I can't even run the defragmenter without a  
quick failure.
I didn't payed big attention to the logs and quickly decided to buy a  
SATA Sil PCI card to get rid of the Nvidia SATA HBA.

- Now the problem.

Yesterday, however, the MPEG2 streams hanged for a few tens of  
seconds just as usual. But there were no disk failure. The array was  
still in good shape, although dmesg showed the same "ata[56]:  
Resetting port", "SCSI errors" etc. fuss.
However this was new in dmesg: "anticipatory: forced dispatching is  
broken (nr_sorted=1), please report this". Got 4 identical in a row.
Maybe managing 8 block devices queues under load with the  
anticipatory scheduler is too much ? I immediately switched to  
deadline on the 8 disks, and I'll see if it it happens again by  
stressing the whole system more and more.
I have no clue if anticipatory is a good choice or definitely not in  
my case, anyone can point some documentation or good advices ?

- How to reproduce.

Here is what I would do:
Harness a small CPU with lots of sata/scsi drives.
Do raid 5 with big block size (1-4Mb) on it.
Make a 50G XFS file system with sunit/swidth options
Trigger bonnie++ with 1G<files<4G and fill the FS to 80-95%, trying  
to achieve 98%+ fragmentation.
Defrag !

- Finally the usual bug report stuff is attached.

Download attachment "cpuinfo" of type "application/octet-stream" (1238 bytes)

View attachment "dmesg-06-01_07-01.txt" of type "text/plain" (30863 bytes)

Download attachment "lspci" of type "application/octet-stream" (25799 bytes)

Download attachment "modules" of type "application/octet-stream" (818 bytes)

Download attachment "scsi" of type "application/octet-stream" (1918 bytes)

Download attachment "ver_linux" of type "application/octet-stream" (925 bytes)


Thanks for your help !

GoM

Powered by blists - more mailing lists