Monday, November 16, 2015


pam_slurm_adopt is a PAM module I wrote that adopts incoming ssh connections into the appropriate Slurm job on the node. The module allows Slurm to control ssh-launched processes as if they were launched under Slurm in the first place and provides for limits enforcement, accounting, and stray process cleanup when a job exits.

Some MPI implementations don't provide for Slurm integration or are not compiled with Slurm support, so the fallback is ssh.  Some code doesn't even use MPI and instead directly calls ssh.  The ideal solution is always to utilize properly-compiled MPI that supports Slurm, but realistically that's not going to happen all the time.  That's where pam_slurm_adopt comes in.

As I see it, this PAM module solves three main problems.  ssh-launched processes now:
  1. Have proper limit enforcement
  2. Will be accounted for in the accounting database
  3. Will be cleaned up when the job exits
We have been using the code in production or more than three weeks now and things work great, except a few minor bugs still need to be fixed to get all the intended benefits.  See the "Inclusion in Slurm" section for details.

Extern Step

The module makes use of the extern step that is created when PrologFlags=contain is set (thanks to SchedMD for adding this feature to support pam_slurm_adopt).

Early versions of pam_slurm_adopt required the use of cgroups for process adoption but no longer do due to the new stepd_add_extern_pid function and RPC.  The extern step is just like any other step, though its accounting information is only recorded if there is some CPU usage in it.  The sshd process handling the incoming connection is placed into that step.  All processes launched from that ssh connection are then tracked, limited, and accounted for by Slurm as part of the extern step.

Identification of the Associated Job

When an ssh connections is made, there is no information available about which Slurm job initiated the ssh connection.  There are also some situations where this determination would be impossible to make anyway, such as when a user on a login node connects to a compute node that is running several jobs from that user.

There are several situations that pam_slurm_adopt has to handle: zero jobs, one job, multiple jobs and we can figure out which job it's a part of, and multiple jobs and we cannot figure out which job it's a part of.  Several parameters control the PAM module's behavior in these situations and are explained below.

The user has zero jobs

By default (action_no_jobs), the module rejects the connection.  It is worth noting that other PAM modules can still allow the connection through depending on your configuration.

The user has only one job

The only reason the user has access to this node is because of the one job, therefore we assume that the incoming connection should be adopted into the job's extern step.  This approach skips other more expensive methods.  If you really want to change this behavior (let me know why so I can find out why it's useful to you), check the source code for the undocumented single_job_skip_rpc parameter.

The user has multiple jobs: Successful RPC

When multiple jobs from the user have allocations on the node, pam_slurm_adopt makes use of the new CallerID RPC that I wrote earlier this year.  The PAM module first determines the source and destination IP addresses and port numbers (IPv4 and IPv6 are supported).  It then uses the CallerID RPC to contact the slurmd that (hopefully) is listening at the source IP address and sends the network connection information to it.

The remote slurmd then uses the network connection information to track down the ssh process that initiated the connection to the other compute node.  Once it locates the process, it then looks up the Slurm job that it is a member of; Slurm's process tracking makes this possible.

After the RPC determines the remote job ID, the PAM module then adopts the ssh process into that job's extern step.

The user has multiple jobs: Unsuccessful RPC

Sometimes the RPC cannot be successful.  If the user initiated the connection from a login node, no slurmd will be running at the IP address that the RPC is made to.  The RPC will either time out or will be rejected by the host OS, preferably the latter.  If the slurmd port on the login node is firewalled in such a way that incoming packets are silently dropped rather than rejected, the RPC must wait for the configurable timeout.

Additionally, there may be a situation where an ssh connection is initiated from a compute node but, for some reason, the slurmd can't associate the source process of the connection with a job.  Maybe you configured the PAM module to allow users access even without jobs, so they aren't part of a job when they ssh again elsewhere.  Maybe you restarted slurmd at the exact moment that the RPC was made from another node.  Maybe there's a bug or you're just unlucky.  For some reason, slurmd just can't figure out the job ID.

The action_unknown parameter defines the (likely controversial) action to take when all else fails.  The default "newest" option is, in my opinion, the best.  I will attempt to convince others why this is the best option by explaining why it's the least bad option.  I will also note that node sharing must be allowed (or you set single_job_skip_rpc=0 for some reason) for you to get to this point.

The "newest" option picks the newest job on the node from the user and adopts the process into that job's extern step. The slurmd doesn't actually have access to the start time or time limit of the jobs, so the most efficient way to query that information is to compare the creation times of the various jobs' step_extern cgroups. The most recent directory mtime is chosen.  For now, it just checks the memory cgroups.  If you would prefer to use a different subsystem, modify the _indeterminate_multiple function.

Unfortunately, "newest" may result in a user running things related to job A but getting adopted into job B then having job B exit before A is finished. This will likely be a very rare event though I'm sure it will happen sometimes.

This replaced an earlier "any" option that just picked the first step it saw based on the output of a particular function call. That resulted in a somewhat random choice of job and often resulted in an older job being chosen.
The "user" option adopts the process into the slurm/uid_$UID/step_extern cgroups.  This uid_$UID cpuset cgroup contains limits that are aggregated for the jobs from that user, but the uid_$UID memory cgroup is currently not set up with an aggregated limit.  This means that the user's CPU usage can be controlled but not memory usage.  Additionally, including the ssh connection's accounting information is impossible since it's not associated with a job.  I don't recall for sure, but I don't think that stray jobs are cleaned up automatically when all user jobs exit the node.  I would not recommend this option at present.
The "user" option was removed when the switch was made from writing directly to cgroups to using stepd_add_extern_pid.  There is no mechanism in stepd_add_extern_pid for writing just to a user cgroup, and it probably wouldn't be too useful anyway.

The "allow" option just allows the process to continue without any adoption.  Limits are not enforced, accounting doesn't take place, and stray processes aren't cleaned up.  This PAM module was worthless for this situation.
The "deny" option does just what it says. This will likely cause many more user support headaches than "newest".

Sometimes a user may have a single job on a node but may sometimes have multiple jobs on a node. If a user logs into a node with only one job, the connection will succeed and be adopted into that job. If the user logs into a node with multiple jobs, the connection will be denied. From a user's point of view, this will appear to be random behavior.

Even worse, there is no good way for the user to specify that a connection should be associated with a particular job. Basically, the user sees random connection failures that can't be circumvented. This seems much worse than the occasional situation with "newest" where a user process gets adopted into the newest job on a node and the newest job exits before the user finishes working with the intended job. Deny will happen every time whereas "newest" will negatively affect someone very rarely.
Is it working?

You can enable more debug output with the PAM module's log_level parameter.  The debug levels are the same as those for SlurmctldDebug.  "debug" is probably sufficient for most debugging and "debug3" gets you slightly more from the pam module itself.  If I recall correctly there is also additional information from other portions of the Slurm code at "debug5" that might be useful for debugging.

Additionally, the /proc/self/cgroup file tells you what cgroups a process is a member of.  If a process is properly launched or adopted into a Slurm job, you should see something like this and maybe more cgroups depending on your configuration:
Those paths are under their respective cgroup root (see /proc/mounts).  The root is controlled by your cgroup.conf's CgroupMountpoint setting, which defaults to /cgroup.  cgroup subsystems are mounted under there, so 2:cpuset:/slurm/uid_1234/job_8804874/step_extern in the example above is /cgroup/cpuset/slurm/uid_1234/job_8804874/step_extern using the default configuration.

Inclusion in Slurm

The necessary code was included in Slurm 15.08.3.  We have been using the latest pam_slurm_adopt code in production starting with 15.08.2, prior to its inclusion in .3.

  • 15.08.3 mostly works and is safe to use
    • Processes are adopted and limited
      • I left off the "s" in "devices" so you need this patch
    • Accounting does not work
    • Stray processes are not cleaned up
  • 15.08.4 reworked to use new stepd_add_extern_pid function
    • Bug 2096: Stray processes are not cleaned up
    • Bug 2097: Accounting not working cpuset cgroup not created
    • Accounting now works
  • 15.08.5 will hopefully be 100% functional when released
    • Bugs 2096 and 2097 sill need to be fixed

Stray process cleanup can be fixed by using epilog to kill off anything in step_extern, such as this (assuming your cgroups are mounted at /cgroup): xargs kill < /cgroup/cpuset/slurm/uid_$UID/job_$SLURM_JOB_ID/step_extern/tasks.  I have not yet implemented or tested that code in epilog (but I will at some point) so test it first.  It may also need to clean up task_$task cgroups under the step_extern.  Once bug 2096 is resolved, this will no longer be necessary.

Hopefully 15.08.5 will fix everything.  I will update this post later.

More Information

More information about the mechanics of the CallerID RPC (REQUEST_NETWORK_CALLERID) can be found in a previous post.  Configuration information can be found with the source code in contribs/pam_slurm_adopt/README.

Let me know if you have any other ideas.

No comments:

Post a Comment

Please leave any comments, questions, or suggestions below. If you find a better approach than what I have documented in my posts, please list that as well. I also enjoy hearing when my posts are beneficial to others.