openvswitch+dpdk 100% cpu usage of ovs-vswitchd

Satish Patel satish.txt at gmail.com
Mon Nov 9 15:11:19 UTC 2020


That would be great, thank you. Let me try to create VF based bonding
inside DPDK and see how it goes.

On Mon, Nov 9, 2020 at 9:30 AM Sean Mooney <smooney at redhat.com> wrote:
>
> On Mon, 2020-11-09 at 09:13 -0500, Satish Patel wrote:
> > Thank Sean,
> >
> > I have Intel NIC
> >
> > [root at infra-lxb-1 ~]# lspci | grep -i eth
> > 06:00.0 Ethernet controller: Intel Corporation 82599 10 Gigabit Dual
> > Port Backplane Connection (rev 01)
> > 06:00.1 Ethernet controller: Intel Corporation 82599 10 Gigabit Dual
> > Port Backplane Connection (rev 01)
> >
> > I was thinking if i can create a couple VF out of SR-IOV interface and
> > on a computer machine i create two bonding interfaces. bond-1 for mgmt
> > and bond-2 for OVS+DPDK then it will solve my all problem related TOR
> > switches redundancy.
> >
> > I don't think we can add VF as an interface in OVS for DPDK.
> you can and if you create the bond on the host first it basically defeets teh reason for using dpdk
> the kernel bond driver will be a bottelneck for dpdk. if you want to bond dpdk interfaces then you should
> create that bond in ovs by adding the two vfs and then creatign an ovs bond.
>
> > On Mon, Nov 9, 2020 at 9:03 AM Sean Mooney <smooney at redhat.com> wrote:
> > >
> > > On Mon, 2020-11-09 at 04:41 +0000, Tony Liu wrote:
> > > > Bonding is a SW feature supported by either kernel or DPDK layer.
> > > > In case of SRIOV, it's not complicated to enable bonding inside VM.
> > > > And it has to be two NICs connecting to two ToRs.
> > > >
> > > > Depending on DPDK implementation, you might be able to use VF.
> > > > Anyways, it's always recommended to have dedicated NIC for SRIOV.
> > > for what its worth melonox do support bondign fo VF on the same card
> > > i have never used it but bonding on the host is possibel for sriov.
> > > im not sure if it works with openstack however but i belvie it does.
> > >
> > > you will have to reach out to mellonox to determin if it is.
> > > most other nic vendors do not support bonding and it may limit other
> > > feature like bandwith based schduling as really you can only list one of the interfaces bandwith
> > > because you cant contol which interface is activly being used.
> > >
> > > >
> > > >
> > > > Thanks!
> > > > Tony
> > > > > -----Original Message-----
> > > > > From: Satish Patel <satish.txt at gmail.com>
> > > > > Sent: Sunday, November 8, 2020 6:51 PM
> > > > > To: Tony Liu <tonyliu0592 at hotmail.com>
> > > > > Cc: Laurent Dumont <laurentfdumont at gmail.com>; OpenStack Discuss
> > > > > <openstack-discuss at lists.openstack.org>
> > > > > Subject: Re: openvswitch+dpdk 100% cpu usage of ovs-vswitchd
> > > > >
> > > > > Thank you tony,
> > > > >
> > > > > We are running openstack cloud with SR-IOV and we are happy with
> > > > > performance but one big issue, it doesn't support bonding on compute
> > > > > nodes, we can do bonding inside VM but that is over complicated to do
> > > > > that level of deployment, without bonding it's always risky if tor
> > > > > switch dies. that is why i started looking into DPDK but look like i hit
> > > > > the wall again because my compute node has only 2 NIC we i can't do
> > > > > bonding while i am connected over same nic. Anyway i will stick with SR-
> > > > > IOV in that case to get more performance and less complexity.
> > > > >
> > > > > On Sun, Nov 8, 2020 at 3:22 PM Tony Liu <tonyliu0592 at hotmail.com> wrote:
> > > > > >
> > > > > > SRIOV gives you the maximum performance, without any SW features
> > > > > > (security group, L3 routing, etc.), because it bypasses SW.
> > > > > > DPDK gives you less performance, with all SW features.
> > > > > >
> > > > > > Depend on the use case, max perf and SW features, you will need to
> > > > > > make a decision.
> > > > > >
> > > > > >
> > > > > > Tony
> > > > > > > -----Original Message-----
> > > > > > > From: Laurent Dumont <laurentfdumont at gmail.com>
> > > > > > > Sent: Sunday, November 8, 2020 9:04 AM
> > > > > > > To: Satish Patel <satish.txt at gmail.com>
> > > > > > > Cc: OpenStack Discuss <openstack-discuss at lists.openstack.org>
> > > > > > > Subject: Re: openvswitch+dpdk 100% cpu usage of ovs-vswitchd
> > > > > > >
> > > > > > > I have limited hands-on experience with both but they don't serve
> > > > > > > the same purpose/have the same implementation. You use SRIOV to
> > > > > > > allow Tenants to access the NIC cards directly and bypass any
> > > > > > > inherent linux- vr/OVS performance limitations. This is key for NFV
> > > > > > > workloads which are expecting large amount of PPS + low latency
> > > > > > > (because they are often just virtualized bare-metal products with
> > > > > > > one real cloud- readiness/architecture ;) ) - This means that a
> > > > > > > Tenant with an SRIOV port can use DPDK + access the NIC through the
> > > > > > > VF which means (in theory) a better performance than OVS+DPDK.
> > > > > > >
> > > > > > > You use ovs-dpdk to increase the performance of OVS based flows (so
> > > > > > > provider networks + vxlan based internal-tenant networks).
> > > > > > >
> > > > > > > On Sun, Nov 8, 2020 at 11:13 AM Satish Patel <satish.txt at gmail.com
> > > > > > > <mailto:satish.txt at gmail.com> > wrote:
> > > > > > >
> > > > > > >
> > > > > > >        Thanks. Just curious then why people directly go for SR-IOV
> > > > > > >       implementation where they get better performance + they can
> > > > > use the
> > > > > > >       same CPU more also. What are the measure advantages or
> > > > > features
> > > > > > >       attracting the community to go with DPDK over SR-IOV?
> > > > > > >
> > > > > > >       On Sun, Nov 8, 2020 at 10:50 AM Laurent Dumont
> > > > > > > <laurentfdumont at gmail.com <mailto:laurentfdumont at gmail.com> > wrote:
> > > > > > >       >
> > > > > > >       > As far as I know, DPDK enabled cores will show 100% usage at
> > > > > > > all times.
> > > > > > >       >
> > > > > > >       > On Sun, Nov 8, 2020 at 9:39 AM Satish Patel
> > > > > > > <satish.txt at gmail.com <mailto:satish.txt at gmail.com> > wrote:
> > > > > > >       >>
> > > > > > >       >> Folks,
> > > > > > >       >>
> > > > > > >       >> Recently i have added come compute nodes in cloud
> > > > > supporting
> > > > > > >       >> openvswitch-dpdk for performance. I am seeing all my PMD
> > > > > > > cpu cores are
> > > > > > >       >> 100% cpu usage on linux top command. It is normal behavior
> > > > > > > from first
> > > > > > >       >> looks. It's very scary to see 400% cpu usage on top. Can
> > > > > someone
> > > > > > >       >> confirm before I assume it's normal and what we can do to
> > > > > > > reduce it if
> > > > > > >       >> it's too high?
> > > > > > >       >>
> > > > > > >
> > > > > >
> > >
> > >
> >
>
>



More information about the openstack-discuss mailing list