Re: Fuse client hangs on doing multithreading IO tests

["lengbo@xxxxxxxxxxxxx" <lengbo@xxxxxxxxxxxxx>]

Hi,

Sorry, the client OS version is CentOS 6.6, not 6.7.

And I find it is not the problem of GlusterFS. It's a kernel problem. futex_wait() may cause deadlock.

https://groups.google.com/forum/#!msg/mechanical-sympathy/QbmpZxp6C64/nMhNjQPTeLEJ 

It just happens on CentOS 6.5. The bug was fixed on kernel 2.6.32-504.16.2. After I upgrade the kernel, then I have not seen this problem until now. I will continue the test to confirm it is solved.

Thanks,

Paul

 

From: paul

Date: 2016-06-25 01:44

To: FNU Raghavendra Manjunath

CC: gluster-devel@xxxxxxxxxxx; gluster-users@xxxxxxxxxxx

Subject: Re: Fuse client hangs on doing multithreading IO tests

Hi, Raghavendra,

The typical file size is 2GB. I remember that the problem existed if I use 10MB file size.

I have attached the logs. The attachment file includes the logs of one client and two servers.

This time the client hung at 16:48 on June 24.  

Thanks,

Paul

 Original Message 

Sender: FNU Raghavendra Manjunath<rabhat@xxxxxxxxxx>

Recipient: 冷波<lengbo@xxxxxxxxxxxxx>

Cc: gluster-devel@xxxxxxxxxxx<gluster-devel@xxxxxxxxxxx>; gluster-users@xxxxxxxxxxx<gluster-users@xxxxxxxxxxx>

Date: Friday, Jun 24, 2016 23:18

Subject: Re: Fuse client hangs on doing multithreading IO tests

Hi,

Any idea how big were the files that were being read?

Can you please attach the logs from all the gluster server and client nodes? (the logs can be found in /var/log/glusterfs)

Also please provide the /var/log/messages from all the server and client nodes.

Regards,

Raghavendra

On Fri, Jun 24, 2016 at 10:32 AM, 冷波 <lengbo@xxxxxxxxxxxxx> wrote:

Hi,

We found a problem when doing traffic tests. We created a replicated volume with two storage nodes (CentOS 6.5). There was one FUSE client (CentOS 6.7) which did multi-threading reads and writes. Most of IOs are reads for big files. All machines used 10Gbe NICs. And the typical read throught was 4-6Gbps (0.5-1.5GB/s). 

After the test ran several minutes, the test program hung. The throughput suddenly dropped to zero. Then there was no traffic any more. If we ran df, df would hang, too. But we could still read or write the volume from other clients.

We tried several GlusterFS version from 3.7.5 to 3.8.0. Each version had this problem. We also tried to restore default GlusterFS options, but the problem still existed.

The GlusterFS version was 3.7.11 for the following stacks.

This was the stack of dd when hanging:

[<ffffffffa046d211>] wait_answer_interruptible+0x81/0xc0 [fuse]

[<ffffffffa046d42b>] __fuse_request_send+0x1db/0x2b0 [fuse]

[<ffffffffa046d512>] fuse_request_send+0x12/0x20 [fuse]

[<ffffffffa0477d4a>] fuse_statfs+0xda/0x150 [fuse]

[<ffffffff811c2b64>] statfs_by_dentry+0x74/0xa0

[<ffffffff811c2c9b>] vfs_statfs+0x1b/0xb0

[<ffffffff811c2e97>] user_statfs+0x47/0xb0

[<ffffffff811c2f9a>] sys_statfs+0x2a/0x50

[<ffffffff8100b072>] system_call_fastpath+0x16/0x1b

[<ffffffffffffffff>] 0xffffffffffffffff

This was the stack of gluster:

[<ffffffff810b226a>] futex_wait_queue_me+0xba/0xf0

[<ffffffff810b33a0>] futex_wait+0x1c0/0x310

[<ffffffff810b4c91>] do_futex+0x121/0xae0

[<ffffffff810b56cb>] sys_futex+0x7b/0x170

[<ffffffff8100b072>] system_call_fastpath+0x16/0x1b

[<ffffffffffffffff>] 0xffffffffffffffff

This was the stack of the test program:

[<ffffffff810a3f74>] hrtimer_nanosleep+0xc4/0x180

[<ffffffff810a409e>] sys_nanosleep+0x6e/0x80

[<ffffffff8100b072>] system_call_fastpath+0x16/0x1b

[<ffffffffffffffff>] 0xffffffffffffffff

Any clue?

Thanks,

Paul

_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-devel

_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-devel