Re: High ping times on interconnect for Cisco UCS servers....
Date: Fri, 1 Mar 2013 18:14:12 -0800
Message-ID: <CADsdiQgJGGTc=2orp_HO_YY6RTs-bE+fK4UpZx4+WT28CHxyvg_at_mail.gmail.com>
Ping:
"the surprising result was that for nearby hosts (<< 1 ms) the
> times ping gives you are more to do with ping code and CPU thread
> scheduling rather than network latency."
>
http://mail.opensolaris.org/pipermail/dtrace-discuss/2008-March/005680.html
I don't particularly trust ping for low latency testing ie under 1ms. If ping is over 1ms regularly or dropping packets then that's another situation.
I typically see 300us back to back on machines in the lab. Occasionally I see 150us but never looked exactly into why. When I run network performance tests the test run as fast as would be expected.
For network testing I've been using 3 tools
- netio - simple good push button throughput test
- netperf - more nobs and whistles, a bit overwhelming
- ttcp - java based net test tool
I want to try out iperf but just haven't gotten around to it.
For netperf and ttcp, I've written some shell scripts to drive load. My
ttcp scripts are a little further along capturing netstat info but ttcp has
a lot of java stack overhead that influences timings.
Using netperf is much lighter.
With ttcp all you need, beside java, is on github
https://github.com/khailey/ttcp
With netperf, you will probably need to compile the client end if you want to get the stats I'm parsing in the shell script. I put the basic binaries and shell scripts up on
https://github.com/khailey/netperf
Also the netperf.sh script doesn't capture the netstat data which is useful for looking at retransmissions. The code from ttcp.sh could be plugged directly into netperf.sh to capture netstats, I just haven't done it yet
- Kyle Hailey http://dboptimizer.com
On Thu, Feb 28, 2013 at 5:27 PM, Bobak, Mark <Mark.Bobak_at_proquest.com>wrote:
> Hi,
> Does anyone have any experience with Cisco UCS servers?
>
> I've just been handed my first batch of them, and I'm getting ready to
> install a two-node RAC. These servers are replacement for HP DL360s. I
> currently have a two-node RAC running on the DL360s. The ping times on
> this RAC, with bonded 1GigE NICs, are around 0.070 ms. On the new UCS
> servers (10GigE "vNICs"), my ping times are in the neighborhood of 0.170.
> That's about 2.4 times slower...
>
> I fully expected *faster* ping times for 10GigE interface....
>
> So, has anyone seen something like this before? Any thoughts? Ideas?
>
> FYI, I also have a second two-node UCS server RAC that I'm getting ready
> to set up, and it seems to have the same problem.
>
> Any ideas, thoughts, or suggestions, would be appreciated.
>
> Thanks,
>
> -Mark
>
> --
> http://www.freelists.org/webpage/oracle-l
>
>
>
-- http://www.freelists.org/webpage/oracle-lReceived on Sat Mar 02 2013 - 03:14:12 CET