[hipl-users] Re: performance measurement

  • From: Miika Komu <miika@xxxxxx>
  • To: HIPL Users <hipl-users@xxxxxxxxxxxxx>
  • Date: Fri, 20 Feb 2004 17:25:32 +0200 (EET)

On Fri, 20 Feb 2004, Mika Kousa wrote:

> On Fri, 20 Feb 2004, Simon Schuetz wrote:
>
> -> has anybody used a tool to measure TCP performance over HIP?
> -> I recompiled an IPv6 version of netperf and linked it to the supplied
> -> libinet6. It's using HIP communication, but it doesn't seem to work
> -> properly. Any other tools used yet?
>
> We have done some small network throughput tests with ttcp (with IPv6
> support from http://www.netcore.fi/pekkas/linux/ipv6/). I personally have
> no experience on netperf.

As I recall, I had trouble with making the IPv6 support in netperf with
HIP. Or was it that netperf did not support IPv6... I don't remember
anymore.

Notice that there are a couple of things to be noticed when measure HIP
performance:

* TCP timeouts for 3 seconds if the base exchange is executed at the same
  time. This happens because the kernel drops the first SYN instead of
  caching it and using timers. The SYN is sent only when the timeout is
  finished and TCP retransmits. There is a quick workaround for this by
  changing the timeout value constant in the kernel, but that is
  honestly a kludge.
* The first UDP packet is always lost if base exchange is executed
  simultaneously.
* If you measure REA on the same link, you may experience some timeouts
  due to IPv6 dupplicate address detection.

We have already done some performance measurements but they are not
currently public.

-- 
Miika Komu              miika@xxxxxx          http://www.iki.fi/miika/

Other related posts: