Poor Adaptec 6911A(tulip) performance w/options=14

[Kai Peter Ford]

We have a little 4-node cluster which with Adaptec 6911A NICs (DEC 21143).

We've run variations on the following netperf and netpipe commands:

netperf -H <remote host> -- -D -m <msg size> -s <socket size> -S <socket size>
NPtcp -s -t -h <remote host> -l ... -u ... -b <socket size>

with msg size from a few Bytes to MB's and socket size up to 64K.

With "options tulip options=14" in /etc/conf.modules (MII 100baseTX-FD), throughput increases 
with increasing socket size up to about 55Mb/s at 8K socket size, then plummets to just a few 
Mb/s for socket sizes greater than 8K.  

With "... options=13" (MII 100baseTX), throughput increases with socket size to over 80 Mb/s 
at 64K socket size, e.g., there is no performance fall off for half duplex.

Does anyone know how to get decent full duplex performance with these cards?


Details:

Asus P2B-D 440BX, 350MHz PII (2 in master node), 
256MB (512 in master), 
Adaptec 6911A/TXC (DEC 21143-PA tulip) NICs.
Foundry FastIron Workgroup Fast Ethernet switch.

* We have tried the following Asus/6911A/tulip variations: 
Foundry FastIron switch, or cross-over cable
Linux 2.0.35-1 (RH5.1), 2.0.36-SMP (patched RH5.1, 1 node only), 2.0.36-0.7 (RH5.2), 
2.0.36-3 (upgraded RH5.2).
tulip v0.90, v0.90F,(compiled SMP and no SMP), v0.90Q.
With and without netperf -D option (TCP_NODELAY).
Different message size increments.
tulip options=0, 5, 11, 13, & 14 in /etc/conf.modules (auto, 100baseTx-FD, MII auto, MII 
100baseTx, and MII 100baseTx-FD).
Modular kernel in all cases.

Only options=13 made a significant difference.

--------------------------------------------------------
Kai Peter Ford (kaiford@cascadesci.com)
Cascade Scientific Software, Inc.
"Software Engineering Services for Scientific Applications"
WWW:   http://www.halcyon.com/chilcoat/cascade/cascade.html
Email: kaiford@cascadesci.com
Phone: (425) 489-9320
Fax:   (425) 489-0955
--------------------------------------------------------