After considerable time spent researching this problem, and with the
help of Hal Stern (stern@sunne.East.Sun.COM), my Sparcstation 2
servers are no longer dropping packets in large NFS tranfers.
The problem was diagnosed by turning on debugging for the le interface
on the NFS server with:
echo "ledebug/W 1" | adb -k -w /vmunix /dev/mem
This produced debugging messages:
le0: out of tmds - packet dropped
This occurred about 5-7 times per 4 meg file transfer.
The fix was to increase the number of transmit descriptors in the
Lance Ethernet driver:
In the file /sys/sunif/le_conf.c, in the stanza under
#ifdef NFSSERVER, change the value of le_low_ntmdp2 from 6 to 7.
This increases the number of transmit descriptors from 64 to 128.
Config and rebuild your kernel, reboot.
I noticed this in a heterogenous environment. We are also apparently
seeing some additional timeouts on the receive side of our RS6000's,
above and beyond this problem.
Ahhhh, the joys of pushing Ethernet to the limit.
Thanks again to Hal Stern, who provided the answer to this one.
-Mahlon
-------------------------------------------------------------
Mahlon Stacy Internet: mcs@mayo.edu
Mayo Foundation
Rochester, MN 55905 Minnesota Regional Network
(507) 284-4558 Amateur: KF0AW
-------------------------------------------------------------
This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:06:19 CDT