Xen 4.3 NUMA Aware Scheduling

Note: Originally posted on blog.xenproject.org

Background and Motivation

This document provides an update on NUMA Aware scheduling in Xen 4.3. You can find other articles on NUMA in the NUMA categort.

Long story  short, they say how NUMA is becoming more and more common and that, therefore, it is very important to:

1. achieve a good initial placement, when creating a new VM;
2. have a solution that is both flexible and effective enough to take advantage of that placement during the whole VM lifetime.

The former, basically, means: <<When starting a new Virtual Machine, to which NUMA node should I "associate" it with?>>. The latter is more about: <<How hard should the VM be associated to that NUMA node? Could it, perhaps temporarily, run elsewhere?>>.

NUMA Placement and Scheduling

So, here's the situation: automatic initial placement has been included in Xen 4.2, inside libxl. This means, when a VM is created (of course, if that happens through libxl) a set of heuristics decide on which NUMA node his memory has to be allocated, and the vCPUs of the VM are statically pinned to the pCPUs of such node.

On the other hand, NUMA aware scheduling has included in Xen 4.3. This mean, instead of being statically pinned, the vCPUs of the VM will strongly prefer to run on the pCPUs of the NUMA node, but they can run somewhere else as well... And this is what this status report is all about.

Performance Numbers?

Sure thing! Benchmarks similar to the ones already described in the past have been performed. More specifically, directly from the cover letter of the Xen 4.3 implementation. Here's what has been done:

I ran the following benchmarks (again):
* SpecJBB is all about throughput, so pinning is likely the ideal
  solution.
* Sysbench-memory is the time it takes for writing a fixed amount
  of memory (and then it is the throughput that is measured). What
  we expect is locality to be important, but at the same time the
  potential imbalances due to pinning could have a say in it.
* LMBench-proc is the time it takes for a process to fork a fixed
  number of children. This is much more about latency than
  throughput, with locality of memory accesses playing a smaller
  role and, again, imbalances due to pinning being a potential
  issue.

This all happened on a 2 node host, where 2 to 10 VMs (2 vCPUs and 960 RAM each) were executing the various benchmarks concurrently. Here they are the results:

 ----------------------------------------------------
 | SpecJBB2005, throughput (the higher the better)  |
 ----------------------------------------------------
 | #VMs | No affinity |  Pinning  | NUMA scheduling |
 |    2 |  43318.613  | 49715.158 |    49822.545    |
 |    6 |  29587.838  | 33560.944 |    33739.412    |
 |   10 |  19223.962  | 21860.794 |    20089.602    |
 ----------------------------------------------------
 | Sysbench memory, throughput (the higher the better)
 ----------------------------------------------------
 | #VMs | No affinity |  Pinning  | NUMA scheduling |
 |    2 |  469.37667  | 534.03167 |    555.09500    |
 |    6 |  411.45056  | 437.02333 |    463.53389    |
 |   10 |  292.79400  | 309.63800 |    305.55167    |
 ----------------------------------------------------
 | LMBench proc, latency (the lower the better)     |
 ----------------------------------------------------
 | #VMs | No affinity |  Pinning  | NUMA scheduling |
 ----------------------------------------------------
 |    2 |  788.06613  | 753.78508 |    750.07010    |
 |    6 |  986.44955  | 1076.7447 |    900.21504    |
 |   10 |  1211.2434  | 1371.6014 |    1285.5947    |
 ----------------------------------------------------

Which, reasoning in terms of %-performance increase/decrease, means NUMA aware scheduling does as follows, as compared to no-affinity at all and to static pinning:

     ----------------------------------
     | SpecJBB2005 (throughput)       |
     ----------------------------------
     | #VMs | No affinity |  Pinning  |
     |    2 |   +13.05%   |  +0.21%   |
     |    6 |   +12.30%   |  +0.53%   |
     |   10 |    +4.31%   |  -8.82%   |
     ----------------------------------
     | Sysbench memory (throughput)   |
     ----------------------------------
     | #VMs | No affinity |  Pinning  |
     |    2 |   +15.44%   |  +3.79%   |
     |    6 |   +11.24%   |  +5.72%   |
     |   10 |    +4.18%   |  -1.34%   |
     ----------------------------------
     | LMBench proc (latency)         |
     | NOTICE: -x.xx% = GOOD here     |
     ----------------------------------
     | #VMs | No affinity |  Pinning  |
     ----------------------------------
     |    2 |    -5.66%   |  -0.50%   |
     |    6 |    -9.58%   | -19.61%   |
     |   10 |    +5.78%   |  -6.69%   |
     ----------------------------------

The tables show how, when not in overload (where overload='more vCPUs than pCPUs'), NUMA scheduling is the absolute best. In fact, not only it does a lot better than no-pinning on throughput biased benchmarks, as well as a lot better than pinning on latency biased benchmarks (especially with 6 VMs), it also equals or beats both under adverse circumstances (adverse to NUMA scheduling, i.e., beats/equals pinning in throughput benchmarks, and beats/equals no-affinity on the latency benchmark).

When the system is overloaded, NUMA scheduling scores in the middle, as it could have been expected. It must also be noticed that, when it brings benefits, they are not as huge as in the non-overloaded case. However, this only means that there is still room for more optimization, right?  In some more details, the current way a pCPU is selected for a vCPU that is waking-up, couples particularly bad with the new concept of NUMA node affinity. Changing this is not trivial, because it involves rearranging some locks inside the scheduler code, but is already being worked-on. Anyway, even with what we have right now, we are overloading the test box by 20% here (without counting Dom0 vCPUs!) and still seeing improvements, which is definitely not bad!

XXX

NUMA Scheduling

Suppose we have a VM with all its memory allocated on NODE#0 and NODE#2 of our NUMA host. Of course, the best thing to do would be to pin the VM’s vCPUs on the pCPUs related to the two nodes. However, pinning is quite unflexible: what if those pCPUs get very busy while there are completely idle pCPUs on other nodes? It will depend on the workload, but it is not hard to imagine that having some chance to run –even if on a remote node– would be better than not running at all. It would therefore be preferable to give the scheduler some hints about where a VM’s vCPUs should be executed. It then can try at its best to honor these requests of ours, but not at the cost of subverting its own algorithm. From now on, we’ll call this hinting mechanism node affinity (don’t confuse it with CPU affinity, which is about to static CPU pinning).

As said, the experimental patchset introduces also the support for node affinity aware scheduling by means of this changeset: [sched_credit: Let the scheduler know about `node affinity`. As of now, it is all very simple, and it only happens for the credit1 scheduling plugin of the Xen hypervisor. However, looking at some early performance measurements seems promising.

Looking at the results, attempts to suggest the scheduler the preferred node for a VM seem to be the righ direction to go (For the interested, complete results set is here). The various curves in the graph below represents the throughput achieved on one of the VMs, more specifically the one that is being, respectively:

• scheduled without any pinning or affinity, i.e., (cpus="all" in the VM config file, red line, also called default in this article),
• created and pinned on NODE#0, so that all its memory accesses are local (green line),
• scheduled with node affinity only to NODE#0 (no pinning) as per what is introduced by the patch (blue line).

What the plot actually shows is the percent increase of each configuration with respect to the worst possible case (i.e., when all memory access are remote). This means tweaking node affinity increases performance by ~12% to ~18% from the worst case. Also, it lets us gain up to ~8% performance as compared to unpinned behavior, doing particularly well as load increases. However, although it gets quite close to the green line (which is the best case), there is still probably some performance bits to squeeze from it.

Combining Placement and Scheduling

So, if we do both things, i.e.:

• we introduce some form of automatic placement logic in xl. This means assigning a "node affinity" to a domain at creation time and asking Xen to stick to it;
• we tweak the scheduler so that it will strongly prefer running a domain on the node(s) it has affinity with, but not in a strict manner as it is for pinning.

The various lines have the same meaning described in the Xen NUMA Introduction article.

Here it is what the benchmarks tells. In all the graphs here, the light blue line is the interesting one, as it is representative of the case when VM#1 has its affinity set to NODE#0. The most interesting among the various plots is probably this one below:

We can see the "node affinity" curve managing in getting quite closer to the best case, especially under high system load (4 to 8 VMs). It can't be called as perfect yet, as some more consideration needs to be given to the not-so-loaded cases, but it is a start. If you feel like wanting to help with testing, benchmarking, fixing or whatever... Please, jump in!