Good Morning Andreas,

 

Very valuable insight one gets with this command.

I see quite a few of these requests coming in and completing ever few seconds. So this confirms something is still writing to it based on the opc 4 = OST_WRITE (I just found out from one of your older posts J )

And this is not just one client, many clients are writing to it. I have quite a number of jobs running for weeks now.

I also looked at the clients and since they are all running several jobs there are over 2500 files open from /luster, and I could not find any deleted files from lsof’s output, except on local file system.

 

My guess is that there are active jobs in our cluster that still have a file handle open with objects on OST’s before I deactivated them? Is this possible? Any thoughts?

 

…..

[root@diskarray6 ~]# lctl get_param ost.OSS.ost_io.req_history | grep "opc 4"

[root@diskarray6 ~]# lctl get_param ost.OSS.ost_io.req_history | grep "opc 4"

1222327203:10.1.1.59@tcp:12345-10.1.11.54@tcp:x1421713825905502:448:Complete:1382101014:0s(-6s) opc 4

1222327204:10.1.1.59@tcp:12345-10.1.13.17@tcp:x1421714339953255:448:Complete:1382101014:0s(-6s) opc 4

1222327205:10.1.1.59@tcp:12345-10.1.13.32@tcp:x1431400659252398:448:Complete:1382101014:0s(-6s) opc 4

[root@diskarray6 ~]# lctl get_param ost.OSS.ost_io.req_history | grep "opc 4"

[root@diskarray6 ~]# lctl get_param ost.OSS.ost_io.req_history | grep "opc 4"

1222327210:10.1.1.59@tcp:12345-10.1.11.58@tcp:x1421713794727740:448:Complete:1382101019:0s(-6s) opc 4

1222327211:10.1.1.59@tcp:12345-10.1.11.77@tcp:x1421713757657163:448:Complete:1382101019:0s(-6s) opc 4

1222327212:10.1.1.59@tcp:12345-10.1.13.43@tcp:x1421714872793796:448:Complete:1382101019:0s(-6s) opc 4

1222327213:10.1.1.59@tcp:12345-10.1.11.49@tcp:x1421713827559437:448:Complete:1382101019:0s(-6s) opc 4

1222327214:10.1.1.59@tcp:12345-10.1.13.34@tcp:x1421714086296893:448:Complete:1382101020:0s(-6s) opc 4

…..

 

Best Regards,
Amit

 

From: hpdd-discuss-bounces@lists.01.org [mailto:hpdd-discuss-bounces@lists.01.org] On Behalf Of Pascal Decourt
Sent: Friday, October 18, 2013 5:35 AM
To: Dilger, Andreas
Cc: hpdd-discuss@lists.01.org
Subject: Re: [HPDD-discuss] OST deactivated but still objects being written to it?

 

Thx a lot Andreas.

Pascal.

On 18/10/2013 11:46, Dilger, Andreas wrote:

On 2013/10/18 1:53 AM, "Pascal Decourt" <pdecourt@sgi.com> wrote:
Very interesting.
I run the lctl command and here is the output:
 
 
65838502:10.149.254.2@o2ib:12345-10.149.1.171@o2ib:x1448244599539665:448:I
nterpret:1382081942:-1382081942s(-1382081952s) opc 4
 
The size of the history buffer is controlled via the
"req_buffer_history_max"
tunable.  By default it only has a few entries in it.
 
For my understanding, could you please give more details on the field of
this line ?
 
Fields are separated by colons.
 
1st field: a sequence counter, so it is possible to determine if a
record has already been processed.
 
2nd field: target NID (i.e. server)
 
3rd field: source NID (i.e. client)
 
4th field: client RPC XID number (per-client unique identifier)
 
5th field: RPC request buffer size
 
6th field: state of the RPC processing
 
7th field: time (unix epoch) when RPC arrived
 
8th field: total processing time (seconds)
 
9th field: time before client would have timed out request
 
10th field: RPC opcode number (see lustre_idl.h)
 
This is all described in the user manual if you search for req_history:
 
https://wiki.hpdd.intel.com/display/PUB/Documentation
 
There is also a tool "lustre_req_history" that processes these files a bit.
 
Cheers, Andreas
 
On 18/10/2013 07:14, Kumar, Amit wrote:
 
 
Awesome Andreas sounds promising, I will give this a shot in the
morning....
Regards 
Amit
 
Sent from my iPhone
 
On Oct 17, 2013, at 6:19 PM, "Dilger, Andreas" <andreas.dilger@intel.com>
<mailto:andreas.dilger@intel.com> wrote:
 
On 2013/10/14 11:56 AM, "Kumar, Amit" <ahkumar@mail.smu.edu>
<mailto:ahkumar@mail.smu.edu> wrote:
I Just saw your message. Even though I have it deactivated,  something is
still writing significantly. Hope you can help me understand this better.
Even though new objects are not assigned can the disk usage increase for
the existing objects?
 
On the other hand, my migration script is not able to locate any files on
the OST to move, which is odd. One thing I know is some files names are
very odd, like may be 20-30 spaces in a file of length about at least
100 characters. Some users program is fanatic in file name creation and I
have no control over it.
 
Hope you can shed some more light. I am planning on running fsck to
identify any issues with the ost.
 
 
I don't think fsck is needed.  It seems like you have some process on a
client that
is writing to an open-unlinked file on those OSTs.
 
To find out which client, run on the affected OSS:
# lctl get_param ost.OSS.ost_io.req_history | grep "opc 4"
ost.OSS.ost_io.req_history=
1431689:192.168.20.1@tcp:12345-192.168.20.159@tcp:x1448908035689984:488:Co
m
plete:1382050012:0s(-43s) opc 4
1431690:192.168.20.1@tcp:12345-192.168.20.159@tcp:x1448908035689988:488:Co
m
plete:1382050012:0s(-43s) opc 4
 
This will print out all clients (NID "192.168.20.159@tcp" in this case)
that
are writing.  On the client(s), use "lsof | grep /lustre/mountpoint" to
see
what processes have files open on the filesystem (probably marked
"(deleted)")
and kill those processes.
 
This should immediately free up all of the space on those OSTs.
 
Cheers, Andreas 
 
-----Original Message-----
From: Nico Budewitz [mailto:Nico.Budewitz@aei.mpg.de]
Sent: Thursday, October 10, 2013 3:14 PM
To: Kumar, Amit
Cc: hpdd-discuss@lists.01.org
Subject: Re: [HPDD-discuss] OST deactivated but still objects being
written to it?
 
Hi,
 
temp. deactivated OSTs via 'lctl --device devno deactivate' will be
marked as inactive. No new objects are assigned to the deactivated OST,
but reads and writes of existing objects are still no problem.
 
Lustre_Manual Chapter: Removing an OST from the File System Hope that
helps
- --
"la lykken gro, som gresset bak do"
Nico Budewitz
High Performance Computing
Max-Planck-Institute for Gravitational Physics /
Albert-Einstein-Institute Am Muehlenberg 1, 14476 Golm
Tel.: +49 (0)331 567 7364 Fax: +49 (0)331 567 7284
http://supercomputers.aei.mpg.de
 
On Oct 10, 2013, at 10:01 PM, "Kumar, Amit" <ahkumar@mail.smu.edu>
<mailto:ahkumar@mail.smu.edu> wrote:
 
Dear All,
 
I have deactivated an OST yet I see the size of the OST is increasing
constantly?
 
Any thoughts what could be causing this?
 
Thank you,
Amit
 
 
Cheers, Andreas

 

--
Pascal Decourt
System Support Engineer Europe | SGI France
+33 1 34 88 80 31 | +33 6 80 45 24 77
pdecourt@sgi.com

SGI

Silicon Graphics SA
21 rue Albert Calmette
78350 Jouy en Josas
  RCS Versailles B 344 515 291