SUMMARY: ce Gigaswift driver fails to set infinite_burst=1 on second Gigabit card

From: Alek O. Komarnitsky (N-CSC) <alek_at_ast.lmco.com>
Date: Mon May 12 2003 - 14:24:28 EDT
In a summary I just sent out about FTP's to localhost being slower on 
Solaris2.9 versus 2.8, I mentioned a minor bug in the ce Gigaswift driver
where it fails to set infinite_burst for a SECOND ce card - here are the
details on that. 

I don't fully understand what infinite_burst actually does,
but basically it seems to be a way of really cranking up 
the speed of your PCI card/ce driver.

Sun patch 112817-04 has this bug report/fix in it:
   4717385 infinite_burst should default to 1 on Schizo platforms
and there is some discussion buried in the Gigaswift documentation
on SunSolve that indicates you DO want this on for Sun280R's.

We initially had a lower rev installed, and infinite_burst was
NOT set at all - this manifested itself as slower FTP performance
between two "close" GigE machines - instead of ~60 MBytes/second,
we saw about 30 MBytes/second.


On our Sun280R's with TWO ce cards (on copper, on fiber), 
we installed the patch (112817-07 actually) and upon doing a:
   ndd /dev/ce infinite_burst
this was set to 1 on /dev/ce1 (the fiber), but was set to 0
on /dev/ce0 (you have to do a "ndd -set /dev/ce instance #"
to query whatever device you are looking at) and we still saw
diminished performance on the ce0 interface.

The fix was fairly straightforward - add this line to /kernel/drv/ce.conf
   infinite_burst=1;
and reboot ... both interfaces came up with infinite_burst=1 and
performance was good.

As an aside, in testing, we tried setting infinite_burst=0 in ce.conf;
thinking that would set it for BOTH interfaces - it does NOT ... so
something in the ce driver code overrides it ... but as indicated above,
my guess is it doesn't correctly handle the case of multiple ce cards.

alek
http://www.komar.org/


P.S. I'll try to get someone here to "officially" submit this 
to Sun as a bug, but I can't promise that will happen ...
so if someone on the list wants to followup, I've outlined it
pretty well - my guess is a fairly easy code fix.
_______________________________________________
sunmanagers mailing list
sunmanagers@sunmanagers.org
http://www.sunmanagers.org/mailman/listinfo/sunmanagers
Received on Mon May 12 14:24:23 2003

This archive was generated by hypermail 2.1.8 : Thu Mar 03 2016 - 06:43:10 EST