From: "Avi Cohen (A)" <avi.cohen@huawei.com>
To: "Wiles, Keith" <keith.wiles@intel.com>
Cc: "users@dpdk.org" <users@dpdk.org>
Subject: Re: [dpdk-users] OVS vs OVS-DPDK
Date: Wed, 24 May 2017 13:51:36 +0000 [thread overview]
Message-ID: <B84047ECBD981D4B93EAE5A6245AA361013BCD16@FRAEML521-MBX.china.huawei.com> (raw)
In-Reply-To: <365623D9-223D-4A37-ACB7-73599B4E163C@intel.com>
Thanks Keith for your reply
I found out that the bottleneck are the VMs and not the OVS/OVS-DPDK running in the host.
VMs on both setup are unaware to OVS/OVS-DPDK and use their linux IP-stack.
I found that the performance (e.g. throughput) between VMa - OVS-DPDK - network - OVS-DPDK - VMb is much better than with standard OVS.
I use vhost-user virtio for the OVS-DPDK setup to connect to VM , and vhost-net for the standard OVS
The reasons for standard OVS poor performance can be for example:
1. number of packet copies in the path NIC - OVS - OS-guest-virtio - Application on guest
2. interrupt upon receiving a packet
3. # of context-switch / VM-exit
etc..
I didn't see any info regarding these potential reasons on the docs.
Best Regards
avi
> -----Original Message-----
> From: Wiles, Keith [mailto:keith.wiles@intel.com]
> Sent: Wednesday, 24 May, 2017 4:23 PM
> To: Avi Cohen (A)
> Cc: users@dpdk.org
> Subject: Re: [dpdk-users] OVS vs OVS-DPDK
>
>
> > On May 24, 2017, at 3:29 AM, Avi Cohen (A) <avi.cohen@huawei.com>
> wrote:
> >
> > Hello
> > Let me ask it in a different way:
> > I want to understand the reasons for the differences in performance between
> OVS-DPDK and standard OVS My setup is: ovs/ovs-dpdk is running @ host
> communicating with a VM
> >
> > OVS-DPDK
> > 1. packet is received via physical port to the device.
> >
> > 2.DMA transfer to mempools on huge-pages allocated by dpdk-ovs - in
> user-space.
> >
> > 3. OVS-DPDK copies this packet to the shared-vring of the associated guest
> (shared between ovs-dpdk userspace process and guest)
> >
> > 4. guest OS copies the packet to userspace application on VM .
> >
> > Standard OVS
> >
> > 1. packet is received via physical port to the device.
> >
> > 2.packet is processed by the OVS and transferred to a virtio device connected
> to the VM - whar are the additional overhead here ? QEMU processing -
> translation , VM exit ?? other ?
> >
> > 3. guest OS copies the packet to userspace application on VM .
> >
> >
> > Question: what are the additional overhead in the standard OVS that cause
> to poor performance related to the OVS-DPDK setup ?
> > I'm not talking about the PMD improvements (OVS-DPDK) running on the
> host - but on overhead in the VM context in the standard OVS setup
>
> The primary reasons are OVS is not using DPDK and OVS is using the Linux
> kernel as well :-)
>
> >
> > Best Regards
> > avi
>
> Regards,
> Keith
next prev parent reply other threads:[~2017-05-24 13:51 UTC|newest]
Thread overview: 4+ messages / expand[flat|nested] mbox.gz Atom feed top
2017-05-24 8:29 Avi Cohen (A)
2017-05-24 13:23 ` Wiles, Keith
2017-05-24 13:51 ` Avi Cohen (A) [this message]
2017-05-25 9:03 ` Avi Cohen (A)
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=B84047ECBD981D4B93EAE5A6245AA361013BCD16@FRAEML521-MBX.china.huawei.com \
--to=avi.cohen@huawei.com \
--cc=keith.wiles@intel.com \
--cc=users@dpdk.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).