[PVFS-users] Bonnie benchmarks

[bill]

Rob and Kevin,

I am not a Unix power junky just yet, but what I do know is a little about
networks.  My toolkit includes Network Application Perfomance and modeling.
A wonderful little application is called Optimal Networks. Take some sniffer
samples and run them through and it pretty much graphs out the problem. More
often than not the "network" is fingered but is actually the least point of
resistence.  Kevin, if you want we can take this off line and report back to
the group later on? Either you could mail me the traces and network
description or I live about and hour away and would enjoy coming over. Just
a few quick question came to my head while reading.

1) Is it a shared network? -hub
2) Is it switched? What type? Cut through? Store and forward?
3) Who else do you share bandwidth with?
4) Are all ports and NIC's hard coded with speed and duplex?
5) Cheapy NIC's?

Take care all

Bill Northrup


> Kevin,
>
> Heh...yeah, we've seen similar behavior before.  I have tried to track
> this down with little success.  Perhaps it is a TCP-related issue (slow
> start or odd packetization)?  Is the behavior consistent over multiple
> runs for you?  We've seen some cases where it is consistently like this,
> and others where it is not.
>
> I'd love to track down what is causing this :).
>
> Rob
>
> On Thu, 20 Jul 2000, Kevin Makles wrote:
>
> > I've been running a few benchmarks to test out PVFS and found some
strange
> > results when I ran the Bonnie Benchmark.  The strange results only occur
in
> > the  Block write benchmark.  Here are the results from a 4 I/O node
system
> > with 100Mbit ethernet interconnect using version 1.4.4 of PVFS.
> >
> >  KB out     input
> >   3  2483   4237
> >   4  5334   4369
> >   5  3681   5654
> >   6  2973   6238
> >   7  4736   6736
> >   8  4705   6356
> >   9  5030   7470
> >  10   563   7577
> >  11  4672   7800
> >  14  3904   8417
> >  15  5611   8659
> >  16   820   8795
> >  17  4434   8789
> >  18  4702   8981
> >  30  2183   9828
> >  31  4924   9871
> >  32 10022   9915
> >  33  2420   9968
> >  34  4584   9904
> >  64  3243  10596
> > 128  5048  10996
> >
> > Where KB is the value of Chunk in the Bonnie source and out and input
columns
> > are in KB/s
> > --
> > Kevin Makles
> > NERSC Future Technologies Group
> > Lawerence Berkeley Laboratory
>
> _______________________________________________
> PVFS-users mailing list
> PVFS-users@www.beowulf-underground.org
> http://beowulf-underground.org/mailman/listinfo/pvfs-users