NUMA and torque/Moab

May 9, 2013
hpc numa torque moab sysadmin rants work

TL;DR Support for NUMA systems in torque/Moab breaks existing means of specifying shared memory jobs and limits scheduling flexibility in heterogeneous compute environments.

Torque NUMA implementation

For torque NUMA, you manually define your NUMA nodes within a system, which gives you some flexibility for how topology information is exposed to the queue system. The provided mom_gencfg utility defaults to defining a node per CPU sockets, but you can manually craft your mom.layout file to aggregate sockets together into larger NUMA nodes if desired.

While you run a single pbs_mom on the SMP, you’ll see that torque tracks one “compute node” per NUMA node in your system. To illustrate, if you have a 4-socket system with the hostname “smallsmp”, you may see the following:

> pbsnodes -l
smallsmp-0 offline
smallsmp-1 offline
smallsmp-2 offline
smallsmp-3 offline

There are some interesting side effects to this approach. For one, yes you can offline individual sockets. The other side of that is you have to offline all of the sockets if you’re draining the system for maintenance. I think this also highlights why you may not want to do this on anything except your big SMPs (e.g., you have 1000 dual socket boxes, you suddenly see 2000 every time you type pbsnodes).

Submitting shared memory jobs

In our environment, we use two ways to submit shared memory jobs given our Moab configuration. We typically do “qsub -l nodes=1:ppn=N” to request N cores on a single shared memory node. I’ve used tpn as well since I found the interpretation of ppn a little unintuitive if you use something other than “nodes=1”. For distributed memory jobs, we use “nodes=N” and Moab is free to distribute or pack cores across systems (we aim to maximize throughput over single job performance and let jobs share nodes).

Moab NUMA limitations

In theory, Moab supports torque NUMA integration. But it is currently unable to grok that it’s dealing with an SMP and enforces unnecessary limitations if you specify a shared memory job using ppn/tpn. For example, if I have 8 cores per socket in my SMP and I configure a NUMA node per socket, then I can only run jobs requesting up to “ppn=8”. If I request any more than that, Moab won’t schedule them because it is unable to aggregate NUMA nodes within the same system to run a shared memory job submitted in this manner. Torque/Moab really treat your NUMA SMP as a distributed memory cluster.

The way they address this is by telling you to submit shared memory jobs like you would distributed memory jobs, but tack on some extra flags to constrain it to one of your SMPs (e.g., “-l nodes=N,flags=sharedmem”). This is less than ideal for our HPC environment, and Adaptive has been unable to recommend a satisfactory workaround for us since we raised the issue with them 4 months ago.

Our environment and NUMA scheduling limitations

Our HPC environment consists of two clusters managed by a single torque/Moab instance. One is a commodity cluster that’s a mix of 8-core and 16-core nodes with between 24 and 128GB of memory per node. The other is a small cluster of SMPs with the largest having 256 cores (8 cores per socket) and 4TB of memory in a single system image.

When possible, we install new software packages on both clusters. We then write queue submit scripts for our users that abstracts away the details of the queue system and let’s them focus on the app and resources that they need for a particular job (e.g., cores, memory, runtime). By setting it up this way and managing it all with a single queue system instance, we are able to abstract away where the job runs as long as it’s installed on both clusters and allows jobs to run on the first cluster that has the required resources available, easily load balancing across the two clusters to keep our cores busy.

We unfortunately didn’t realize the scheduling limitations imposed by torque/Moab NUMA until after deploying it on our new SMP. Previously, we’d written custom utilities that handled processor core allocation to jobs in a NUMA topology-aware way and configured a single large node entry in the queue system config. We’d hoped to move away from this toward using more standard/supported tools, but didn’t have the time/resources to adequately test the new NUMA support to discover its shortcomings.

How we submit shared memory jobs now

Here’s what we have implemented currently in our submit scripts based on how many cores a user requests for a shared memory app installed on both clusters:

  • For up to 8 cores, “-l nodes=1:ppn=N,partition=commodity:smp”
  • For 9-16 cores, “-l nodes=1:ppn=N,partition=commodity”
  • For more than 16 cores, “-l nodes=N,partition=smp”

Small jobs can take advantage of flexible scheduling across both clusters. Large jobs can only run on the SMP due to hardware constraints. We were faced with a choice for medium-sized jobs. We could either submit them so they’d run on the SMP only or submit them so they’d only run on the larger commodity nodes. We opted for the latter since we have more capacity on the commodity cluster.

Next steps

During our next maintenance window, I intend to try reworking our NUMA config to aggregate both sockets on a node board into a single NUMA node so we’d have 16 cores instead of 8 per NUMA node. Assuming that works as advertised, we should be able to submit small to medium shared memory jobs so they’d run on either cluster… at least until we deploy commodity nodes with more than 16 cores, which may happen as soon as this fall with the release of Ivy Bridge.

I’ve had several discussions with high level folks at Adaptive about this (and other issues related to the 4.1 series). I’m hopeful that I’ve pressed the issue strongly enough with them that they make it a priority on their development roadmap to better address these limitations, but it’s anyone’s guess when/if we’ll see a release that provides a satisfactory resolution to my complaints.