charmm scalability on 2.4 kernels

[Eray Ozkural (exa)]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Greetings Josip,

On Tuesday 08 January 2002 17:41, Josip Loncaric wrote:
> Actually, Linux kernel 2.4 incorporates an improved yet very different
> TCP stack where my TCP fixes do not help much in my usual test case
> (point-to-point streaming of small messages).  However, Steve's
> scalability problems with the stock 2.4 TCP are very interesting, since
> they involve many machines, i.e. a completely different communication
> pattern from my usual point-to-point tests.
>

I have some results regarding 2.4.x. Actually, the 2.2.x was more consistent 
in terms of latency. I'm attaching the output of mpptest performance
test suite on our 32 node beowulf system.

The kernel version is 2.4.14 on all nodes. Network hardware is Intel 82557 
Ethernet PRO 100 (rev 8), and 3COM SuperStack II 3900 100BaseTX switch.

These are for point to point, bisection bandwidth and broadcast tests 
included in the suite. I am very concerned by this situation as it might have 
impact on our research in fine-grained algorithms. (But so far it doesn't 
seem to have)

To plot the results:

gnuplot pt2pt.mpl
gnuplot bisect.mpl
gnuplot bcast.mpl

As you can see there are unexpected spikes in the plots. I don't know what to 
attribute to them but AFAIK there were no other parallel applications running 
at the time, and I am certain that I took multiple runs because of the 
inconsistency of the results.

This is not caused by hardware level problems since in 2.2.x (without any 
patch) there was a monotonous increase in latency with growing message size.

> Old 2.2+fix combination was pretty efficient at aggregating small
> messages into larger TCP packets before sending, in fact better than
> stock 2.4.  Packet aggregation is something that depends on delicate
> timing of congestion control events on both sender and receiver; this is
> very sensitive to the application's communication pattern.  Perhaps a
> 2.4 TCP fix would need to be developed after all...
>

In the bisect and bcast plots, there seem to be fewer irregularities but I 
have no results now to compare them against what 2.2.x did. So it is hard to 
say whether 2.4.x actually improves upon stock 2.2.x.

Sincerely,

- -- 
Eray Ozkural (exa) <erayo at cs.bilkent.edu.tr>
Comp. Sci. Dept., Bilkent University, Ankara
www: http://www.cs.bilkent.edu.tr/~erayo
GPG public key fingerprint: 360C 852F 88B0 A745 F31B  EA0F 7C07 AE16 874D 539C
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.0.6 (GNU/Linux)
Comment: For info see http://www.gnupg.org

iD8DBQE8PdUKfAeuFodNU5wRAnFrAJ9F+Sibp9nP0P+jsrTbmNljB/1WMgCcC1DS
gwlSbAr9lOu9msgiAcOUnoU=
=zn69
-----END PGP SIGNATURE-----
-------------- next part --------------
A non-text attachment was scrubbed...
Name: beo32-2.4-perf.tar.gz
Type: application/x-gzip
Size: 4299 bytes
Desc: mpptest performance results
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20020110/b90225c6/attachment.bin>