Hi,
last week I've rewritten my btl-tcp component to improve several
aspects, mainly no oversubscription of interfaces.
I now have:
- the MCA parameter btl_tcp_disable_family={4|6}
to force the use of a special address family at runtime
- a working include/exclude list for interfaces (broken with
the old design)
- the ability to use different address families for different
directions of communication, so
,--------------. ,---------------.
| Host A | ---- IPv4 --------> | Host B |
| RFC1918/NAT | | IPv4 public |
| IPv6 | <--- IPv6 -------- | IPv6 |
`--------------' `---------------'
And probably something more ;) Yesterday, Thomas Peiselt,
Christian Kauhaus and me (from the Cluster- & Metacomputing Group at the
University of Jena) also did a first benchmark and we're proud to come
up with very good news:
IPv6 is not much slower than IPv4.
Latency and bandwidth utilisation behave comparably, resulting
in 111.17 MB/s with IPv4 versus 109.52 MB/s with IPv6 running IMB 2.3
PingPong over 1GigEthernet (in other words: IPv6 delivered 98.5% of
the IPv4 bandwidth).
I expect to confirm this tendency (less than 2% performance loss)
by detailed benchmarking next week.
Thomas is going to improve address family selection, so whenever
it's possible, use IPv4 and by this squeeze out the last 1.5%
of bandwidth ;) (my implementation tends to prefer IPv6, but
I guess he'll fix it)
Christian is very interested in testing. We have automatic builds,
a cluster with IPv4/IPv6 and a traffic light turning from green
to red whenever a build/test fails, so he's looking forward
to provide testing facilities to the OMPI project, but I guess
he'll raise his questions (mainly with respect to MTT) in a
separate mail.
I'm going to upload the new btl-tcp component to tmp/adi-ipv6
within the next days, it's currently only a branch in our local
subversion which I'd like to merge before.
--
mail: a...@thur.de http://adi.thur.de PGP: v2-key via keyserver
Mit leerem Kopf nickt es sich leichter.
