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-authenti...
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 = 26161449Check 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.