Subject: Re: poor network performance on de0's on 21164's.
To: None <tls@rek.tjls.com>
From: Andrew Gallatin <gallatin@cs.duke.edu>
List: port-alpha
Date: 07/11/1999 19:46:04
Thor Lancelot Simon writes:
 > On Sat, Jul 10, 1999 at 06:34:02PM -0400, Paul H. Anderson wrote:
 > > 
 > > (Mike - mts@off.to and I work together, so I'm speaking up for a
 > > second...)
 > > 
 > > In our setup, no apparent combination of switch, hub, direct wire, full,
 > > half duplex, Tulip or Intel card appears to work properly with our 164LX
 > > boxes.
 > > 
 > > It seems that lots of people report partial or complete success with other
 > > boxes, but after a tremendous amount of time trying many, many different
 > > combinations (all with 164LX 533MHz alphas), we are simply unable to get
 > > good symmetric performance, or even adequate performance.
 > 
 > Are you aware that the chipset used on the 164LX boards has at least
 > one bug which can substantially impact PCI DMA performance?  There's a
 > workaround, but it's not clear that it restores performance, just correct
 > operation.

PCI DMA performance on the LX/SX is mediocre for DMA reads, but it should
certainly be able to keep up with a 100Mb nic.  Using a Myrinet PCI
benchmark which measures host-pci DMA performance, we see about
70MB/sec for DMA reads and about 128MB/sec for DMA writes using 32-bit
cards on our 600Mhz 164lx.  For 64 bit cards, those numbers go to
98MB/sec & 235MB/sec respectively.  

And since NetBSD appears to use the 1GB direct map for all PCI DMA,
the broken pyxis scatter/gather TLB is a non-issue.

 > This is one good reason to buy the older PC164 boards and _not_ the SX
 > or LX boards.  IIRC Jason Thorpe looked into this at some point in the
 > past and determined that the SX boards had had a hardware workaround
 > applied by DEC but that it wasn't clear that the LX boards had...

I believe that the PC164 boards use the cia2 chipset.  While it is not
buggy in terms of correctness, the cia2 chipset also has performance
quirks.  If its anything like the cia chipset that we have in our
AlphaStation 500/266 & 600/266 boxes, I'd expect to see the exact
opposite performance characteristics.  Eg, good performance for DMA
reads & mediocre performance for DMA writes.

The only really good, non-server class PCI chipset I've seen from DEC
is the tsunami (in use on xp1000, ds10, ds20, etc).  But those are,
well, rather expensive..

Drew
------------------------------------------------------------------------------
Andrew Gallatin, Sr Systems Programmer	http://www.cs.duke.edu/~gallatin
Duke University				Email: gallatin@cs.duke.edu
Department of Computer Science		Phone: (919) 660-6590