[slurm-users] Munge log-file fills up the file system to 100%

[Ole Holm Nielsen via slurm-users]

We have some new AMD EPYC compute nodes with 96 cores/node running
RockyLinux 8.9. We've had a number of incidents where the Munge log-file
/var/log/munge/munged.log suddenly fills up the root file system, after a
while to 100% (tens of GBs), and the node eventually comes to a grinding
halt! Wiping munged.log and restarting the node works around the issue.

I've tried to track down the symptoms and this is what I found:

1. In munged.log there are infinitely many lines filling up the disk:

2024-04-11 09:59:29 +0200 Info: Suspended new connections while
processing backlog

2. The slurmd is not getting any responses from munged, even though we run
"munged --num-threads 10". The slurmd.log displays errors like:

[2024-04-12T02:05:45.001] error: If munged is up, restart with
--num-threads=10
[2024-04-12T02:05:45.001] error: Munge encode failed: Failed to
connect to "/var/run/munge/munge.socket.2": Resource temporarily unavailable
[2024-04-12T02:05:45.001] error: slurm_buffers_pack_msg:
auth_g_create: RESPONSE_ACCT_GATHER_UPDATE has authentication error

3. The /var/log/messages displays the errors from slurmd as well as
NetworkManager saying "Too many open files in system".
The telltale syslog entry seems to be:

Apr 12 02:05:48 e009 kernel: VFS: file-max limit 65536 reached

where the limit is confirmed in /proc/sys/fs/file-max.

We have never before seen any such errors from Munge. The error may
perhaps be triggered by certain user codes (possibly star-ccm+) that might
be opening a lot more files on the 96-core nodes than on nodes with a
lower core count.

My workaround has been to edit the line in /etc/sysctl.conf:

fs.file-max = 131072

and update settings by "sysctl -p". We haven't seen any of the Munge
errors since!

The version of Munge in RockyLinux 8.9 is 0.5.13, but there is a newer
version in https://github.com/dun/munge/releases/tag/munge-0.5.16
I can't figure out if 0.5.16 has a fix for the issue seen here?

Questions: Have other sites seen the present Munge issue as well? Are
there any good recommendations for setting the fs.file-max parameter on
Slurm compute nodes?

Thanks for sharing your insights,
Ole

--
Ole Holm Nielsen
PhD, Senior HPC Officer
Department of Physics, Technical University of Denmark

--
slurm-users mailing list -- slurm...@lists.schedmd.com
To unsubscribe send an email to slurm-us...@lists.schedmd.com

[Jeffrey T Frey via slurm-users]

https://github.com/dun/munge/issues/94


The NEWS file claims this was fixed in 0.5.15. Since your log doesn't show the additional strerror() output you're definitely running an older version, correct?


If you go on one of the affected nodes and do an `lsof -p <munged-pid>` I'm betting you'll find a long list of open file descriptors — that would explain the "Too many open files" situation _and_ indicate that this is something other than external memory pressure or open file limits on the process.

[Ole Holm Nielsen via slurm-users]

Hi Jeffrey,

Thanks a lot for the information:

On 4/15/24 15:40, Jeffrey T Frey wrote:
> https://github.com/dun/munge/issues/94

I hadn't seen issue #94 before, and it seems to be relevant to our
problem. It's probably a good idea to upgrade munge beyond what's
supplied by EL8/EL9. We can build the latest 0.5.16 RPMs by:

wget
https://github.com/dun/munge/releases/download/munge-0.5.16/munge-0.5.16.tar.xz
rpmbuild -ta munge-0.5.16.tar.xz

I've updated my Slurm Wiki page
https://wiki.fysik.dtu.dk/Niflheim_system/Slurm_installation/#munge-authentication-service
accordingly now.

> The NEWS file claims this was fixed in 0.5.15. Since your log doesn't show the additional strerror() output you're definitely running an older version, correct?

Correct, we run munge 0.5.13 as supplied by EL8 (RockyLinux 8.9).

> If you go on one of the affected nodes and do an `lsof -p <munged-pid>` I'm betting you'll find a long list of open file descriptors — that would explain the "Too many open files" situation _and_ indicate that this is something other than external memory pressure or open file limits on the process.

Actually, munged is normally working without too many open files as seen
by "lsof -p `pidof munged`" over the entire partition, where the munged
open file count is only 29. I currently don't have any broken nodes with
a full file system that I can examine.

Therefore I believe that the root cause of the present issue is user
applications opening a lot of files on our 96-core nodes, and we need to
increase fs.file-max. And upgrade munge as well to avoid the log file
growing without bounds.

I'd still like to know if anyone has good recommendations for setting the

fs.file-max parameter on Slurm compute nodes?

Thanks,
Ole

Department of Physics, Technical University of Denmark,
Fysikvej Building 309, DK-2800 Kongens Lyngby, Denmark
E-mail: Ole.H....@fysik.dtu.dk
Homepage: http://dcwww.fysik.dtu.dk/~ohnielse/
Mobile: (+45) 5180 1620

[Bjørn-Helge Mevik via slurm-users]

Ole Holm Nielsen via slurm-users <slurm...@lists.schedmd.com> writes:

> Therefore I believe that the root cause of the present issue is user
> applications opening a lot of files on our 96-core nodes, and we need
> to increase fs.file-max.

You could also set a limit per user, for instance in
/etc/security/limits.d/. Then users would be blocked from opening
unreasonably many files. One could use this to find which applications
are responsible, and try to get them fixed.

--
Regards,
Bjørn-Helge Mevik, dr. scient,
Department for Research Computing, University of Oslo

[Ole Holm Nielsen via slurm-users]

Hi Bjørn-Helge,

On 4/16/24 12:08, Bjørn-Helge Mevik via slurm-users wrote:
> Ole Holm Nielsen via slurm-users <slurm...@lists.schedmd.com> writes:
>
>> Therefore I believe that the root cause of the present issue is user
>> applications opening a lot of files on our 96-core nodes, and we need
>> to increase fs.file-max.
>
> You could also set a limit per user, for instance in
> /etc/security/limits.d/. Then users would be blocked from opening
> unreasonably many files. One could use this to find which applications
> are responsible, and try to get them fixed.

That sounds interesting, but which limit might affect the kernel's
fs.file-max? For example, a user already has a narrow limit:

ulimit -n
1024

whereas the permitted number of user processes is a lot higher:

ulimit -u
3092846

I'm not sure how the number 3092846 got set, since it's not defined in
/etc/security/limits.conf. The "ulimit -u" varies quite a bit among our
compute nodes, so which dynamic service might affect the limits?

Perhaps there is a recommendation for defining nproc in
/etc/security/limits.conf on compute nodes?

Thanks,
Ole

[Bjørn-Helge Mevik via slurm-users]

Ole Holm Nielsen <Ole.H....@fysik.dtu.dk> writes:

> Hi Bjørn-Helge,

>
> That sounds interesting, but which limit might affect the kernel's
> fs.file-max? For example, a user already has a narrow limit:
>
> ulimit -n
> 1024

AFAIK, the fs.file-max limit is a node-wide limit, whereas "ulimit -n"
is per user.

Now that I think of it, fs.file-max of 65536 seems *very* low. On our
CentOS-7-based clusters, we have in the order of tens of millions, and
on our Rocky 9 based clusters, we have 9223372036854775807(!)

Also a per-user limit of 1024 seems low to me; I think we have in the
order of 200K files per user on most clusters.

But if you have ulimit -n == 1024, then no user should be able to hit
the fs.file-max limit, even if it is 65536. (Technically, 96 jobs from
96 users each trying to open 1024 files would do it, though.)

> whereas the permitted number of user processes is a lot higher:
>
> ulimit -u
> 3092846

I guess any process will have a few open files, which I believe count
against the ulimit -n for each user (and fs.file-max).

> I'm not sure how the number 3092846 got set, since it's not defined in
> /etc/security/limits.conf. The "ulimit -u" varies quite a bit among
> our compute nodes, so which dynamic service might affect the limits?

There is a vague thing in my head saying that I've looked for this
before, and found that the default value dependened on the size of the
RAM of the machine. But the vague thing might of course be lying to
me. :)

--
Bjørn-Helge

[Jeffrey T Frey via slurm-users]

AFAIK, the fs.file-max limit is a node-wide limit, whereas "ulimit -n"
is per user.

The ulimit is a frontend to rusage limits, which are per-process restrictions (not per-user).

The fs.file-max is the kernel's limit on how many file descriptors can be open in aggregate.  You'd have to edit that with sysctl:

$ sysctl fs.file-max

fs.file-max = 26161449

Check in e.g. /etc/sysctl.conf or /etc/sysctl.d if you have an alternative limit versus the default.

But if you have ulimit -n == 1024, then no user should be able to hit
the fs.file-max limit, even if it is 65536.  (Technically, 96 jobs from
96 users each trying to open 1024 files would do it, though.)

Naturally, since the ulimit is per-process the equating of core count with the multiplier isn't valid.  It also assumes Slurm isn't setup to oversubscribe CPU resources :-)

I'm not sure how the number 3092846 got set, since it's not defined in
/etc/security/limits.conf.  The "ulimit -u" varies quite a bit among
our compute nodes, so which dynamic service might affect the limits?

If the 1024 is a soft limit, you may have users who are raising it to arbitrary values themselves, for example.  Especially as 1024 is somewhat low for the more naively-written data science Python code I see on our systems.  If Slurm is configured to propagate submission shell ulimits to the runtime environment and you allow submission from a variety of nodes/systems you could be seeing myriad limits reconstituted on the compute node despite the /etc/security/limits.conf settings.

The main question needing an answer is _what_ process(es) are opening all the files on your systems that are faltering.  It's very likely to be user jobs' opening all of them, I was just hoping to also rule out any bug in munged.  Since you're upgrading munged, you'll now get the errno associated with the backlog and can confirm EMFILE vs. ENFILE vs. ENOMEM.

[Jason Simms via slurm-users]

As a related point, for this reason I mount /var/log separately from /. Ask me how I learned that lesson...

Jason

--
slurm-users mailing list -- slurm...@lists.schedmd.com
To unsubscribe send an email to slurm-us...@lists.schedmd.com

--

Jason L. Simms, Ph.D., M.P.H.

Instructor, Department of Languages & Literary Studies

Lafayette College
Pardee Hall | One Pardee Dr, 4th Fl | Easton, PA 18042
Office: Pardee 405

[Bjørn-Helge Mevik via slurm-users]

Jeffrey T Frey via slurm-users <slurm...@lists.schedmd.com> writes:

>> AFAIK, the fs.file-max limit is a node-wide limit, whereas "ulimit -n"
>> is per user.
>
> The ulimit is a frontend to rusage limits, which are per-process restrictions (not per-user).

You are right; I sit corrected. :)

(Except for number of procs and number of pending signals, according to
"man setrlimit".)

Then 1024 might not be so low for ulimit -n after all.

--
Regard,

[Ole Holm Nielsen via slurm-users]

I looked at some of our busy 96-core nodes where users are currently
running the STAR-CCM+ CFD software.

One job runs on 4 96-core nodes. I'm amazed that each STAR-CCM+ process
has opened almost 1000 open files, for example:

$ lsof -p 440938 | wc -l
950

and that on this node the user has almost 95000 open files:

$ lsof -u <username> | wc -l
94606

So it's no wonder that 65536 open files would have been exhausted, and
that my current limit is just barely sufficient:

$ sysctl fs.file-max
fs.file-max = 131072

As an experiment I lowered the max number of files on a node:

$ sysctl fs.file-max=32768

and immediately the syslog display error messages:

Apr 18 10:54:11 e033 kernel: VFS: file-max limit 32768 reached

Munged (version 0.5.16) logged a lot of errors:

2024-04-18 10:54:33 +0200 Info: Failed to accept connection: Too many
open files in system
2024-04-18 10:55:34 +0200 Info: Failed to accept connection: Too many
open files in system
2024-04-18 10:56:35 +0200 Info: Failed to accept connection: Too many
open files in system
2024-04-18 10:57:22 +0200 Info: Encode retry #1 for client UID=0 GID=0
2024-04-18 10:57:22 +0200 Info: Failed to send message: Broken pipe
(many lines deleted)

Slurmd also logged some errors:

[2024-04-18T10:57:22.070] error: slurm_send_node_msg: [(null)]
slurm_bufs_sendto(msg_type=RESPONSE_ACCT_GATHER_UPDATE) failed: Unexpected
missing socket error
[2024-04-18T10:57:22.080] error: slurm_send_node_msg: [(null)]
slurm_bufs_sendto(msg_type=RESPONSE_PING_SLURMD) failed: Unexpected
missing socket error
[2024-04-18T10:57:22.080] error: slurm_send_node_msg: [(null)]
slurm_bufs_sendto(msg_type=RESPONSE_PING_SLURMD) failed: Unexpected
missing socket error


The node became completely non-responsive until I restored fs.file-max=131072.

Conclusions:

1. Munge should be upgraded to 0.5.15 or later to avoid the munged.log
filling up the disk. I summarize this in the Wiki page
https://wiki.fysik.dtu.dk/Niflheim_system/Slurm_installation/#munge-authentication-service

2. We still need some heuristics for determining sufficient values for the
kernel's fs.file-max limit. I don't understand whether the kernel itself
might set good default values, which we have noticed on some servers and
login nodes.

As Jeffrey points out, there are both soft and hard user limits on the
number of files, and this is what I see for a normal user:

$ ulimit -Sn # Soft limit
1024
$ ulimit -Hn # Hard limit
262144

Maybe the heuristics could be to multiply "ulimit -Hn" by the CPU core
count (if we believe that users will only run 1 process per core). An
extra safety margin would need to be added on top. Or maybe we need
something a lot higher?

Question: Would there be any negative side effect of setting fs.file-max
to a very large number (10s of millions)?

Interestingly, the (possibly outdated) Large Cluster Administration Guide
at https://slurm.schedmd.com/big_sys.html recommends a ridiculously low
number:

> /proc/sys/fs/file-max: The maximum number of concurrently open files. We recommend a limit of at least 32,832.

Thanks for sharing your insights,
Ole

On 4/16/24 14:40, Jeffrey T Frey via slurm-users wrote:
>> AFAIK, the fs.file-max limit is a node-wide limit, whereas "ulimit -n"
>> is per user.
>
> The ulimit is a frontend to rusage limits, which are per-process
> restrictions (not per-user).
>
> The fs.file-max is the kernel's limit on how many file descriptors can be
> open in aggregate.  You'd have to edit that with sysctl:
>
>

> *$ sysctl fs.file-max*

> fs.file-max = 26161449
>
>
>
> Check in e.g. /etc/sysctl.conf or /etc/sysctl.d if you have an alternative
> limit versus the default.
>
>
>
>
>> But if you have ulimit -n == 1024, then no user should be able to hit
>> the fs.file-max limit, even if it is 65536.  (Technically, 96 jobs from
>> 96 users each trying to open 1024 files would do it, though.)
>
> Naturally, since the ulimit is per-process the equating of core count with
> the multiplier isn't valid.  It also assumes Slurm isn't setup to
> oversubscribe CPU resources :-)
>
>
>
>>> I'm not sure how the number 3092846 got set, since it's not defined in
>>> /etc/security/limits.conf.  The "ulimit -u" varies quite a bit among
>>> our compute nodes, so which dynamic service might affect the limits?
>
> If the 1024 is a soft limit, you may have users who are raising it to
> arbitrary values themselves, for example.  Especially as 1024 is somewhat
> low for the more naively-written data science Python code I see on our
> systems.  If Slurm is configured to propagate submission shell ulimits to
> the runtime environment and you allow submission from a variety of
> nodes/systems you could be seeing myriad limits reconstituted on the
> compute node despite the /etc/security/limits.conf settings.
>
>
> The main question needing an answer is _what_ process(es) are opening all
> the files on your systems that are faltering.  It's very likely to be user
> jobs' opening all of them, I was just hoping to also rule out any bug in
> munged.  Since you're upgrading munged, you'll now get the errno
> associated with the backlog and can confirm EMFILE vs. ENFILE vs. ENOMEM.
>
>

--
Ole Holm Nielsen
PhD, Senior HPC Officer
Department of Physics, Technical University of Denmark,
Fysikvej Building 309, DK-2800 Kongens Lyngby, Denmark
E-mail: Ole.H....@fysik.dtu.dk
Homepage: http://dcwww.fysik.dtu.dk/~ohnielse/
Mobile: (+45) 5180 1620

--

[Ole Holm Nielsen via slurm-users]

It turns out that the Slurm job limits are *not* controlled by the normal
/etc/security/limits.conf configuration. Any service running under
Systemd (such as slurmd) has limits defined by Systemd, see [1] and [2].

The limits of processes started by slurmd are defined by LimitXXX in
/usr/lib/systemd/system/slurmd.service, and current Slurm versions have
LimitNOFILE=131072.

I guess that LimitNOFILE is the limit applied to every Slurm job, and that
jobs presumably ought to crash if opening more than LimitNOFILE files?

If this is correct, I think the kernel's fs.file-max ought to be set to
131072 times the maximum possible number of Slurm jobs per node, plus a
safety margin for the OS. Depending on Slurm configuration, fs.file-max
should be set to 131072 times number of CPUs plus some extra margin. For
example, a 96-core node might have fs.file-max set to 100*131072 = 13107200.

Does this make sense?

Best regards,
Ole

[1] "How to set limits for services in RHEL and systemd"
https://access.redhat.com/solutions/1257953
[2]
https://wiki.fysik.dtu.dk/Niflheim_system/Slurm_configuration/#slurmd-systemd-limits

--