[LU-2799] ldlm_cbd: This service may have more threads (192) than the given soft limit (128)

• Status: Resolved
• Project: Lustre
• Affects Version/s: Lustre 2.4.0
• Fix Version/s: Lustre 2.4.0
• Type: Bug
• Priority: Minor
• Reporter: Prakash Surya (Inactive)
• Assignee: Nathaniel Clark
• Resolution: Fixed
• Labels: sequoia, shh
• Severity: 3

Description

I see the following message on the console of one of our Sequoia login nodes:

2013-02-12 15:26:17 Lustre: ldlm_cbd: This service may have more threads (192) than the given soft limit (128)

Is this troubling? What's the net effect of this?

Comments

[Peter Jones]

Nathaniel

Could you please look into this one?

Thanks

Peter

[Nathaniel Clark]

The message is just informational. It will be printed if you specified a number of threads larger than 128 via ptlrpc's ldlm_num_threads module parameter, or if the number of cores is large. See lustre/include/lustre_net.h:250 "example 3" states:

On 64-core machine with 8 partitions we will need LDLM_NTHRS_BASE(24)
threads for each partition to keep service healthy, so total threads
number should be 24 * 8 = 192.

This is not harmful nor should it be worrisome.

[Prakash Surya (Inactive)]

Thanks. It should really be removed in that case. If an administrator can safely ignore the message, it should not make it to the console.

[Prakash Surya (Inactive)]

Unless somebody has a convincing reason why the message should stay, I'm reopening this ticket in the hopes that it is removed.

[Andreas Dilger]

Liang, any reason this message should be kept? Would it be better to limit the number of threads?

Prakash, how many sockets/cores are on this login node?

[Prakash Surya (Inactive)]

48 on this node:

$ cat /proc/cpuinfo | grep proc | wc -l
48

[Nathaniel Clark]

Andreas, Liang,
Should the message be changed to a CWARN instead. If there's an upper limit (even if it's soft) being passed, I would think, it should be logged.

Prakash, Sorry about closing the bug prematurely.

[Andreas Dilger]

I'm more inclined to just quiet the message entirely, i.e. CDEBUG(), since there isn't anything the sysadmin can or should do about it.

I guess at some point we need to look at whether there should be one set of threads running on each of the cores, or if one set of threads per socket is enough?

[Nathaniel Clark]

Patch to make the message a CDEBUG:
http://review.whamcloud.com/5447

[Nathaniel Clark]

Patch has landed

[Prakash Surya (Inactive)]

Thanks!