Re: OT: Network latency benchmarks or tests ? (re NFS)
- From: kyle Hailey <kylelf@xxxxxxxxx>
- To: Jared Still <jkstill@xxxxxxxxx>
- Date: Thu, 12 May 2011 16:45:29 -0700
I would be interested in a tool where I could configure a send size such as
8K,16K,32K and get latency information on these send sizes.
I'm thinking something like netio could be modified to do this, but seems
like someone else must have be interested in these measurements in the past.
The latency times 8K transfers (and multples there of) would be more
relevant to the actual use case of figuring how how much network overhead
there is for transferring database block(s) off of NFS. I'd also expect a
tool that was transferring data would stand a greater chance than ping of
hitting any issues such as packet loss and retransmissions.
Ping has a packet size setting, but when the size is over 1500 bytes it
often seems to hang
Typically, I'd expect ping to be the low end response time but it's been
pointed out that ping can fluctuate a fair bit due to OS scheduling
http://www.mail-archive.com/dtrace-discuss@xxxxxxxxxxxxxxx/msg01284.html
run ping and trace with dtrace
> # ./ping.d -c 'ping -s 192.168.1.108 56 7'
...
> ----192.168.1.108 PING Statistics----
> 7 packets transmitted, 7 packets received, 0% packet loss
> round-trip (ms) min/avg/max/stddev = 0.189/0.430/0.526/0.110
dtrace output:
> ICMP round-trip (ns)
> value ------------- Distribution ------------- count
> 90000 |@@@@@@ 1
> 91000 |@@@@@@@@@@@ 2
> 93000 |@@@@@@ 1
> 98000 |@@@@@@ 1
> 100000 |@@@@@@ 1
> 101000 |@@@@@@ 1
>
> Looking good. Note that the measured times are between 0.090
> and 0.102 ms,
> unlike the summary from ping - which has an average of 0.430 ms!
also checkout "superping.d" in the new dtrace book
These kinds of inaccuracies undermine the measurements for me. One of the
best cases of ping data seems to be, as in smokeping, showing the
variability of network traffic latency and the more variability then, the
reasoning goes, the more issues on the net, but if the OS itself is
introducing much of the variability then that undermines the data and
reasoning.
I'm looking for a tool that would measure the similar kind of latencies as
that one might see in an Oracle database reading off of an NFS mount.
- Kyle
http://dboptimizer.com
On Thu, May 12, 2011 at 1:40 PM, Jared Still <jkstill@xxxxxxxxx> wrote:
> On Thu, May 12, 2011 at 11:27 AM, kyle Hailey <kylelf@xxxxxxxxx> wrote:
>
>> Bit off topic, but I haven't been able to track down a good network
>> latency test for timing of block transfers off of NFS mounts.
>> There are some good network throughput test tools, most notably netio (
>> http://www.ars.de/ars/ars.nsf/docs/netio<http://www.google.com/url?q=http%3A%2F%2Fwww.ars.de%2Fars%2Fars.nsf%2Fdocs%2Fnetio&sa=D&sntz=1&usg=AFrqEzdVNzyguB3g4dOsd2BOx11KCCOz4Q>)
>> which I've been using,
>>
>
> I was about to suggest netio...
>
>
>> Sure there is ping, but ping only goes through the ICMP layer and there is
>> not much in the way of configuration.
>>
>
> Please explain why that's bad, and/or what level you expect to see tested?
> Curios minds want to know.
>
> You might look at this Perl module:
> http://search.cpan.org/~octo/Net-Oping-1.21/lib/Net/Oping.pm
> <http://search.cpan.org/%7Eocto/Net-Oping-1.21/lib/Net/Oping.pm>I have
> not used it, but it may at least have some configuration you are looking
> for.
>
>
> Jared Still
> Certifiable Oracle DBA and Part Time Perl Evangelist
> Oracle Blog: http://jkstill.blogspot.com
> Home Page: http://jaredstill.com
>
Other related posts:
