|
Hello,
We have recently deployed new filesystems running 2.12.5 and our users are hitting a problem with Slurm jobs which write the slurm log files onto the new filesystems.
A bit of context on our environment:
Our main HPC cluster is mounting a subdirectory mount of the wider filesystem via a nodemap. The nodemap settings are as follows:
Of special note, we are setting the 'trusted' property to '1' as we do not do any id-mapping on the main cluster, which are all bound into our central LDAP domain. We are also setting 'admin' to 0, which I understand to be equivalent to 'root_squash' across all nodes in the nodemap.
Consequently we aren't setting the usual settings to configure root_squash, relying on nodemaps for this feature entirely:
The problem we are seeing is that when a user launches a slurm job that will write it's slurm output log files into a directory on this filesystem, then it will sometimes fail to create this log file and abort the job.
I have included the full strace capture from the slurmd daemon when it hits this error, along with a simple slurm job file showing the job definition:
#!/bin/bash
#SBATCH -A SUPPORT-CPU
#SBATCH --output=/rds-d7/project/mjr208/testdir_groups/slurm.out.5
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --time=1:00
echo "STARTING JOB"
pwd
exit 0
The key portion of the strace is here:
Unable to find source-code formatter for language: strace slurmd. Available languages are: actionscript, ada, applescript, bash, c, c#, c++, cpp, css, erlang, go, groovy, haskell, html, java, javascript, js, json, lua, none, nyan, objc, perl, php, python, r, rainbow, ruby, scala, sh, sql, swift, visualbasic, xml, yaml
$ grep 72268 slurm.FAIL.strace | grep -v -E '(futex|tgkill|TKILL)'
...
72268 20:44:49.224728 geteuid() = 0 <0.000005>
72268 20:44:49.224747 umask(002) = 022 <0.000005>
72268 20:44:49.224764 getuid() = 0 <0.000004>
72268 20:44:49.224780 getgid() = 0 <0.000005>
72268 20:44:49.224796 getcwd("/var/spool/slurm/slurmd", 4096) = 24 <0.000005>
72268 20:44:49.224814 getgroups(0, NULL) = 51 <0.000005>
72268 20:44:49.224831 getgroups(51, [815, 901, 902, 904, 905, 910, 1099, 1100, 1101, 8053, 8054, 10573, 14500, 14998, 17501, 41000, 41029, 41037, 41357, 42000, 42006, 42042, 42080, 42100, 42112, 42113, 43000, 430
01, 43003, 43005, 43007, 43009, ...]) = 51 <0.000005>
72268 20:44:49.224870 getuid() = 0 <0.000005>
72268 20:44:49.225163 setresgid(-1, 10573, -1 <unfinished ...>
72268 20:44:49.225175 <... setresgid resumed>) = 0 <0.000008>
72268 20:44:49.225574 setgroups(51, [10573, 50045, 80001, 80004, 80007, 80016, 80019, 43003, 43001, 43005, 43009, 43007, 41357, 80017, 43010, 42042, 42006, 41029, 42080, 41037, 17501, 901, 8054, 905, 14998, 904,
14500, 90022, 815, 902, 1100, 8053, ...] <unfinished ...>
72268 20:44:49.225608 <... setgroups resumed>) = 0 <0.000008>
72268 20:44:49.225895 setresuid(-1, 10573, -1 <unfinished ...>
72268 20:44:49.225907 <... setresuid resumed>) = 0 <0.000008>
72268 20:44:49.225919 open("/dev/null", O_RDONLY) = 10 <0.000006>
72268 20:44:49.225938 fcntl(10, F_SETFD, FD_CLOEXEC) = 0 <0.000005>
72268 20:44:49.225955 open("/rds-d7/project/mjr208/testdir_groups/slurm.out.5", O_WRONLY|O_CREAT|O_TRUNC|O_APPEND, 0666) = -1 EACCES (Permission denied) <0.000785>
So this process is the slurmstepd process running as root forked by slurmd when starting the job, and it appears to be setting it's effective UID and GID to that of the job's user, setting it's groups to that of the user, and then attempting to open the slurm output file at which point it fails.
The permissions on this directory are clearly writeable by the user:
I have also captured debug traces from the MDS when this happens, which are attached in the file:
I have noticed in this file the following line which I think might be significant:
This made me suspect whether the root_squash setting of the nodemap could be to blame here, and sure enough, if I set the 'admin' property to 1 and re-run the slurm job, it runs successfully without error.
To test this I added an additional debugging print line to the __mdd_permission_internal function to just print the value of fsiud and uid before it does the permission check:
and I captured another debug trace from the MDS as before, which is attached in:
- slurm.no-root-squash.SUCCESS.mds.llog
The debug line I added above shows:
So it appears that the 'admin' property on this nodemap is having an effect here that is causing problems for slurm. Are we using this wrong, or is this possibly a bug?
To help isolate this further, I've made a little reproducer C program that tries to copy the sequence of system calls that the slurmstepd daemon is doing before attempting to open the file. Attached are also strace, and debug traces from both the client and MDS for this program, both with and without the nodemap 'admin' property set:
- rds-d7-eacces.c
- reproducer.FAIL.client.llog
- reproducer.FAIL.client.strace
- reproducer.FAIL.mds.llog
- reproducer.no-root-squash.SUCCESS.client.llog
- reproducer.no-root-squash.SUCCESS.mds.llog
At the moment, we are having to ask our users to re-direct their slurm output files to another filesystem, as we don't see this behaviour on our (much) older filesystems. We may also just enable the 'admin' property for now, however I would really like to get to the bottom of why this is failing.
Cheers,
Matt
|
rds-d7-eacces.c