> Why don't you reserve any CPUs for OVS/DPDK or VM usage? All
> published
> performance white papers recommend settings for CPU isolation like
> this
> Mellanox DPDK performance report:
>
>
https://fast.dpdk.org/doc/perf/DPDK_19_08_Mellanox_NIC_performance_report.pdf
<https://mailtrack.io/trace/link/d820a3bc37ae49e92351ef785e3cfb4c21ab5d3c?url=https%3A%2F%2Ffast.dpdk.org%2Fdoc%2Fperf%2FDPDK_19_08_Mellanox_NIC_performance_report.pdf&userId=1840365&signature=eda177b39475fac6>
>
> For their test system:
>
> isolcpus=24-47 intel_idle.max_cstate=0 processor.max_cstate=0
> intel_pstate=disable nohz_full=24-47
> rcu_nocbs=24-47 rcu_nocb_poll default_hugepagesz=1G hugepagesz=1G
> hugepages=64 audit=0
> nosoftlockup
>
> Using the tuned service (CPU partitioning profile) make this process
> easier:
>
> https://tuned-project.org/
<https://mailtrack.io/trace/link/4f3f47457c7163aacfe6bb108c6eff554be9cd4d?url=https%3A%2F%2Ftuned-project.org%2F&userId=1840365&signature=f1504b405d9e880f>
>
Nice tutorial, thanks for sharing. I have checked it and configured our
server like this:
isolcpus=12-19 intel_idle.max_cstate=0 processor.max_cstate=0
nohz_full=12-19 rcu_nocbs=12-19 intel_pstate=disable
default_hugepagesz=1G hugepagesz=1G hugepages=24 audit=0 nosoftlockup
intel_iommu=on iommu=pt rcu_nocb_poll
Even though our servers are NUMA-capable and NUMA-aware, we only have
one CPU installed in one socket.
And one CPU has 20 physical cores (40 threads), so I figured out to use
the "top-most" cores for DPDK/OVS, that's the reason of isolcpus=12-19
You can never have too many cores. On POWER systems I'll sometimes
reserve 76 out of 80 available cores to improve overall throughput.
> >
> > ./usertools/dpdk-devbind.py --status
> > Network devices using kernel driver
> > ===================================
> > 0000:b3:00.0 'MT27800 Family [ConnectX-5] 1017' if=ens2
> drv=mlx5_core
> > unused=igb_uio,vfio-pci
> >
> > Due to the way how Mellanox cards and their driver work, I have not
> bond
> > igb_uio to the interface, however, uio, igb_uio and vfio-pci kernel
> modules
> > are loaded.
> >
> >
> > Relevant part of the VM-config for Qemu/KVM
> > -------------------------------------------
> > <cputune>
> > <shares>4096</shares>
> > <vcpupin vcpu='0' cpuset='4'/>
> > <vcpupin vcpu='1' cpuset='5'/>
>
> Where did you get these CPU mapping values? x86 systems typically
> map
> even-numbered CPUs to one NUMA node and odd-numbered CPUs to a
> different
> NUMA node. You generally want to select CPUs from the same NUMA node
> as
> the mlx5 NIC you're using for DPDK.
>
> You should have at least 4 CPUs in the VM, selected according to the
> NUMA topology of the system.
as per my answer above, our system has no secondary NUMA node, all
mappings are to the same socket/CPU.
>
> Take a look at this bash script written for Red Hat:
>
>
https://github.com/ctrautma/RHEL_NIC_QUALIFICATION/blob/ansible/ansible/get_cpulist.sh
<https://mailtrack.io/trace/link/cf74e0c69acb6d9a348606606825a0320898824a?url=https%3A%2F%2Fgithub.com%2Fctrautma%2FRHEL_NIC_QUALIFICATION%2Fblob%2Fansible%2Fansible%2Fget_cpulist.sh&userId=1840365&signature=9884f8ea2b42399d>
>
> It gives you a good starting reference which CPUs to select for the
> OVS/DPDK and VM configurations on your particular system. Also
> review
> the Ansible script pvp_ovsdpdk.yml, it provides a lot of other
> useful
> steps you might be able to apply to your Debian OS.
>
> > <emulatorpin cpuset='4-5'/>
> > </cputune>
> > <cpu mode='host-model' check='partial'>
> > <model fallback='allow'/>
> > <topology sockets='2' cores='1' threads='1'/>
> > <numa>
> > <cell id='0' cpus='0-1' memory='4194304' unit='KiB'
> > memAccess='shared'/>
> > </numa>
> > </cpu>
> > <interface type='vhostuser'>
> > <mac address='00:00:00:00:00:aa'/>
> > <source type='unix'
> path='/usr/local/var/run/openvswitch/vhostuser'
> > mo$
> > <model type='virtio'/>
> > <driver queues='2'>
> > <host mrg_rxbuf='on'/>
>
> Is there a requirement for mergeable RX buffers? Some PMDs like
> mlx5
> can take advantage of SSE instructions when this is disabled,
> yielding
> better performance.
Good point, there is no requirement, I just took an example config and
though it's necessary for the driver queues setting.
That's how we all learn :-)
>
> > </driver>
> > <address type='pci' domain='0x0000' bus='0x07' slot='0x00'
> > function='0x0'$
> > </interface>
> >
>
> I don't see hugepage usage in the libvirt XML. Something similar to:
>
> <memory unit='KiB'>8388608</memory>
> <currentMemory unit='KiB'>8388608</currentMemory>
> <memoryBacking>
> <hugepages>
> <page size='1048576' unit='KiB' nodeset='0'/>
> </hugepages>
> </memoryBacking>
I did not copy this part of the XML, but we have hugepages configured
properly.
>
>
> > -----------------------------------
> > OVS Start Config
> > -----------------------------------
> > ovs-vsctl --no-wait set Open_vSwitch . other_config:dpdk-init=true
> > ovs-vsctl --no-wait set Open_vSwitch . other_config:dpdk-socket-
> mem="4096,0"
> > ovs-vsctl --no-wait set Open_vSwitch . other_config:dpdk-lcore-
> mask=0xff
> > ovs-vsctl --no-wait set Open_vSwitch . other_config:pmd-cpu-mask=0e
>
> These two masks shouldn't overlap:
>
https://developers.redhat.com/blog/2017/06/28/ovs-dpdk-parameters-dealing-with-multi-numa/
<https://mailtrack.io/trace/link/6c59473e0a8547a9cb80d8f52f9cf5190a0712f6?url=https%3A%2F%2Fdevelopers.redhat.com%2Fblog%2F2017%2F06%2F28%2Fovs-dpdk-parameters-dealing-with-multi-numa%2F&userId=1840365&signature=b01114dc094e5fb9>
>
Thanks, this did really help me understand in which order these
commands should be issued.
So, the problem now is the following.
I did all the changes you shared, and started OVS/DPDK in a proper way
and set these features:
ovs-vsctl --no-wait set Open_vSwitch . other_config:dpdk-socket-
mem="8192,0"
ovs-vsctl --no-wait set Open_vSwitch . other_config:dpdk-lcore-
mask=0x01000
ovs-vsctl --no-wait set Open_vSwitch . other_config:dpdk-init=true
and, finally this:
ovs-vsctl --no-wait set Open_vSwitch . other_config:pmd-cpu-
mask=0x0e000
The documentation you shared say this last one can be even set during
runtime. So, I was playing with it to see there is any change.
I did not start any VM on top of OVS/DPDK, just set up a port forward
rule (in_port=1, actions=output:IN_PORT), since I only have one
physical ports on each mellanox card.
Then, I generated traffic from the other server towards OVS
Using pktsize 64B, the max throughput Pktgen reports is 8Gbps.
In particular, I got these metrics:
Size Sent_pps Recv_pps Recv_Gbps
64B 93M 11M ~8
128B 65M 12.5M ~15
256B 42.5M 12.3M ~27
512B 23.5M 11.9M ~51
1024B 11.9M 10M ~83
1280B 9.6M 8.3M ~86
1500B 8.3M 6.7M ~82
It's quite interesting that for 64B, the pps is less than for greater
sizes. Because PPS should be the practical limitation in throughput,
and according to the packet size we can count the throughput in Gbps.
Looking at 64B performance gives you a sense of the per-packet overhead
associated with the DPDK framework and your application. At 100Gb/s
line rate, 64B frames will arrive every 6.72ns. Since your received PPS
is peaking around 12.5MPPS I'd guess that it's taking about 80ns of CPU
time per frame. I don't know how well OVS scales with additional CPUs,
something to look at.
You don't mention how many different flows you're using in the test.
Don't be surprised as throughput drops when you move from 1,000 flows to
1,000,000 flows.
It's likely that most of your frame loss is due to the NICs RX buffers
overflowing and dropping frames due to back pressure (i.e. DPDK/OVS
can't process packets fast enough). Look the the mlx5's hardware
statistics to verify.
You may be able to improve the performance by increasing the number of
RX queues and RX descriptors per queue, and assigning more lcores to
match the number of queues, allowing the work to be spread more evenly
and reducing buffer overflows. This often works when running testpmd
alone since the app overhead is low but has less effect on OVS
perforamnce. You might consider benchmarking testpmd alone vs OVS/DPDK
to understand the OVS overhead.
Anyway, OVS-DPDK have 3 cores to use, but only one rx queue is assigned
to the port (so, basically --- as `top` also shows --- it is the one-
core performance.
Increasing the number of RX queues/descriptors and assigning a dedicated
lcore to each queue will generally improve performance if your
bottleneck is RX in the PMD.
Increasing the cores did not help, and the performance remained the
same. Is this performance normal for OVS/DPDK?
That's been my experience, though there are other who have more
experience with performance testing OVS. The platform matters. Look
for existing whitepapers and compare your system configuration to theirs
to see what you need to achieve the performance you're looking for.
Dave
