[ale] Question about Virtual Nics and Speed
Brian Mathis
brian.mathis+ale at betteradmin.com
Thu Jun 6 00:47:18 EDT 2013
On Wed, Jun 5, 2013 at 6:56 PM, David Tomaschik <david at systemoverlord.com>wrote:
> On Wed, Jun 5, 2013 at 11:29 AM, Chuck Payne <terrorpup at gmail.com> wrote:
> > Guys,
> >
> > I am currently running a couple Kernel Virtual Machines (KVM) servers,
> > today my db admin told me that he noticed that the speed isn't great
> > on a rsync he was doing. I looked and notice that he was only transfer
> > about 300MB per second. So I started digging into why is that,
> > everything should be at about 1Gig in speed.
> >
> > I told a look network devices on the server with ip
> >
> > ip link show
> <snip>
>
> ip link show doesn't expose the connection speed of your interface, so
> I'm not sure what leads you to believe your vnic is at 500Mb/s. If
> you're talking about the "qlen 500", that's the txqueuelen, or the
> number of packets that the kernel will queue for that interface at any
> given time.
>
> > I am using e1000 as my virtual nics
> >
> > -device e1000
>
> If your guests have a relatively recent kernel, you should be able to
> use virtio-net as your driver instead of e1000. It has significantly
> lower overhead, so if your guests are starving the host CPU, this can
> help some.
>
> >
> > Is there a setting I need to check? Is it control by brctl? Or 500MB
> > the best I will get.
>
> As others have pointed out, 300MB/s is actually not possible over a
> 1gig link. You might be getting 300MB after compression, which would
> be great, or 300Mb, which is not ideal, but also possible.
>
> In all likelihood, your bottleneck is at rotating media somewhere,
> unless you have a fast SAN or RAID array.
>
>
> --
> David Tomaschik
> OpenPGP: 0x5DEA789B
> http://systemoverlord.com
> david at systemoverlord.com
>
To clarify and add to what others have said:
- If you rsync a file that already has some partial files on the other
side, rsync will report it's process of analyzing both sides as part of the
transfer speed. This is the "speed up" that JD mentioned. It seems to
report how fast it's reading off the disk, so it's going to be a lot faster
than your network.
Also, once it runs out of partial data on the disk and starts doing the
network transfer, it will average it all together, which still makes your
network speed seem really fast, if you go by rsync's ETA numbers.
- You should use iperf to measure your network speed. You set it up as a
listener on one side and as a client on the other. It will run some tests
and give you the throughput speed. It's a lightweight thing to run --
nothing big to setup and use.
❧ Brian Mathis
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.ale.org/pipermail/ale/attachments/20130605/1d8f9cd6/attachment.html>
More information about the Ale
mailing list