RS/6000 Multiuser Performance

[Douglas Lee Schales]

In article <78003 at bu.edu.bu.edu> jdh at bu-pub.bu.edu (Jason Heirtzler) writes:
>
>Has anyone else noticed problems with the ethernet controller on the
>model 320?   Some testing I've done seems to indicate that TCP
>throughput is limited to about 400KB/sec, which is quite slow.  Even
>stranger is that if I reduce the size of the individual packets (say
>if the size of the packet changes from 1500 bytes down to 64 bytes) I
>still can't send much more packets/sec than with the larger packet
>size.  The throughput would naturally decrease with smaller packets,
>as you would expect, but you should to be able to send more packets/sec
>if they're smaller.  It's really odd!
>
>It is possible that our controller is still a pre-GA version, and
>we are getting it swapped to a newer one (part 718 1182) to see if
>this helps.

We've noticed that NFS on a 320 here is a real dog.  We attributed it to
the Ethernet card.  A Sparc1+ left the 320 choking on dust when they both
were writing on an NFS partition.  The
application was floating point intensive with about 70k of output.
The Sparc was 2-3 times faster when the output was written to a file on
an NFS partition, while the 320 was
~2 times faster when writing to a local disk.

Our SE is checking into the problem.

Doug.
+---------------------------+
| Douglas Lee Schales       |
| schales at cs.tamu.edu       |
| Dept. of Computer Science |
| Texas A&M University      |
+---------------------------+

--
+---------------------------+
| Douglas Lee Schales       |
| schales at cs.tamu.edu       |
| Dept. of Computer Science |