Open Stack

On Thu, 2020-11-19 at 12:31 +0000, Stephen Finucane wrote:
> On Thu, 2020-11-19 at 12:25 +0000, Stephen Finucane wrote:
> > On Thu, 2020-11-19 at 12:00 +0000, Eyle Brinkhuis wrote:
> > > Hi all,
> > > 
> > > We’re running into an issue with deploying our infrastructure to
> > > run high throughput, low latency workloads.
> > > 
> > > Background:
> > > 
> > > We run Lenovo SR635 systems with an AMD Epyc 7502P processor. In
> > > the BIOS of this system, we are able to define the amount of NUMA
> > > cells per socket (called NPS). We can set 1, 2 or 4. As we run a 2x
> > > 100Gbit/s Mellanox CX5 in this system as well, we use the
> > > preferred-io setting in the BIOS to give preferred io throughput to
> > > the Mellanox CX5.
> > > To make sure we get as high performance as possible, we set the NPS
> > > setting to 1, resulting in a single numa cell with 64 CPU threads
> > > available.
from what data i have personllay seen on this topic this will pessimise your
perfromance and you should be setting it to 4
if you set it to 1 and place the test applicatio on for example cpus 60-64
you will see a performance reduction in comparisone to cpus 4-8

if you enable 4 numa nodes per socket the 3 that do not have the nic will have more or less
teh same performance which should be better then the perfromace when it was on 60-64 
but the one with the nic will have better perfromance which may actully exceed the performance
you see with the vm/applcaiton running on cores 4-8


the preliminary data our perfomance engineers have seen show that some workloads like small packet netwrok io
can see performance improvment of up to 30+% in some workloads (dpdk's testpmd) and 8% improvement in less memory sensitive workloads
setting NPS=4

i know mohammed naser looked into this too for vexhost in the past and was seeing similar effect.
im not sure if you can share your general finding but did you end up with NPS=4 in the end?

> > > 
> > > Next, in Nova (train distribution), we demand huge pages. Hugepages
> > > however, demands a NUMAtopology, but as this is one large NUMA
> > > cell, even with cpu=dedicated or requesting a single numa domain,
> > > we fail:
> > > 
> > > compute03, compute03 fails NUMA topology requirements. No host NUMA
> > > topology while the instance specified one. host_passes
> > > /usr/lib/python3/dist-
> > > packages/nova/scheduler/filters/numa_topology_filter.py:119
> > 
> > Oh, this is interesting. This would suggest that when NPS is
> > configured to 1, the host is presented as a UMA system and libvirt
> > doesn't present topology information for us to parse. That seems odd
> > and goes against how I though newer versions of libvirt worked.
> > 
> > What do you see for when you run e.g.:
> > 
> >  $ virsh capabilities | xmllint --xpath
> > '/capabilities/host/topology' -
> 
> Also, what version of libvirt are you using? Past investigations [1]
> led me to believe that libvirt would now always present a NUMA topology
> for hosts, even if those hosts were in fact UMA.
> 
> [1] https://github.com/openstack/nova/commit/c619c3b5847de85b21ffcbf750c10421d8b7d193

libvirt was broken on amd systems with nps=1 due to a workaround implemented for non x86 architures
https://bugzilla.redhat.com/show_bug.cgi?id=1860231

that should now by adressed but very recently.

> 
> > > Any idea how to counter this? Setting NPS-2 will create two NUMA
> > > domains, but also cut our performance way down.
actully that should improve perfromance based on most benchmarks we have seen
and work we have been doing with amd on this topic.

the data that i have review so far indeicates that the highest memory bandwith and lowest latency
occurs when you expose all the numa nodes on the host by setting NPS to the largest value for you given cpu.
> > 
> > It's worth noting that by setting NP1 to 1, you're already cutting
> > your performance. This makes it look like you've got a single NUMA
> > node but of course, that doesn't change the physical design of the
> > chip and there are still multiple memory controllers, some of which
> > will be slower to access to from certain cores. You're simply mixing
> > best and worst case performance to provide an average. You said you
> > have two SR-IOV NICs. I assume you're bonding these NICs? If not, you
> > could set NPS to 2 and then ensure the NICs are in PCI slots that
> > correspond to different NUMA nodes. You can validate this
> > configuration using tools like 'lstopo' and 'numactl'.
yep setting it to 1 will just disable the reporting of the real numa toplogy
and basically tieing all the memroy contolers in the socket to act as one but that
generally increases latency and decreasess perfromance. it also does not
provide the information need by the kernel or openstack to optimisze.

the main issue we have right now form an openstack point of view is sriov
we support numa affintiy but not not socket affintiy or numa distance
socket affinity is what you want 80% of the time numa distance is much more complex and is what you actully want
but socket affinity is a very good proxy for it. to use sriov with numa guests such as hugepage guests
you have to disable numa affinity for sriov devices in if you have hsots with multiple numa nodes per socket today
if you want to be able to use all cores.

if not all vms will use sriov then you can still use strict/legacy affinity instead of prefered. 
 
> > 
> > Stephen
> > 
> > > Thanks!
> > > 
> > > Regards,
> > > 
> > > Eyle
> > > 
> > 
>