RE: [myrinet] Re: Optimal systems

The comparable numbers for API's UP2000 are:

 DMA rate for 8192 Byte transfers
 	bus_read  (send) = 178 MBytes/s
 	bus_write (recv) = 248 MBytes/s

This is using GM 1.2pre13 running on a 2.2.13 kernel with SMP enabled.
Single and dual processor configurations at 667 MHz & 750 MHz are offered
today.

Tom

On Thursday, April 27, 2000 1:33 AM, Bob Felderman [SMTP:[email protected]]
wrote:
> => I'm impressed that Bob Felderman had hands-on experience with a
> => motherboard that someone just picked out of a hat.
> 
> Just luck... ;-)
> We've been looking for good PCI performance for some time.
> I even bought an SGI VisWS 320 to use the 64-bit PCI.
> Never got linux drivers really working.
> 
> Alphas have always had the best PCI performance since they
> have done 64-bit for some time. 
> 
> => For those of us that might like Myrinet and haven't bought our systems
> => yet, are there any known "best picks" in the Intel or UltraSparc
> => categories?
> 
> Best in 32-bit PCI in the intel world is pertty much everywhere.
> 440BX seems to get peak bandwidth on every machine I have seen.
> 
> Intel Lancewood motherboards are to be avoided for 66MHz PCI since
> they violate the PCI spec and make the 66MHz PCI slot run at 5volts.
> It should be using 3.3V and our card is not happy in that slot.
> 
> I hear Tyan has an Intel 840 board, but I haven't been able to get
> my hands on it.
> 
> So the only intel machine I've seen with "interesting" PCI is the
> Supermicro PIII DME.
> 
> As for Ultrasparcs, I'd avoid them like the plague if you care about 
> PCI performance. Every Ultrasparc I have come across has lousy PCI
performance
> even with 66MHz or 64bit slots.
> 
> I've been trying to collect info from various machines using 
> the PCI64 Myrinet card.
> 
> 
> ES40 is a 4 processor alpha with 2 independent PCI busses, both
> can be running at this rate supposedly. 64bit 33MHz
> ==> ES40-linux/ES40.gm_debug <==  
> DMA rate for 8192 Byte transfers
> 	bus_read  (send) = 215 MBytes/s
> 	bus_write (recv) = 248 MBytes/s
> 
> 
> This is a stock Apple G4. We've been trying to get Apple to address the
issue of the
> bad PCI read performance, but no luck so far. 64bit 33MHz
> ==> G4-linux/g4.myri.com.gm_debug <==
> DMA rate for 4096 Byte transfers
> 	bus_read  (send) = 110 MBytes/s
> 	bus_write (recv) = 240 MBytes/s
> 
> 
> SGI1200 machines with Intel Lancewood boards
> using only the 32bit 33Mhz slots
> ==> bigcache/sgi1202.myri.com.gm_debug <==
> DMA rate for 4096 Byte transfers
> 	bus_read  (send) = 128 MBytes/s
> 	bus_write (recv) = 128 MBytes/s
> 
> Misc compaq machines with 32bit 33MHz slots
> ==> proliant-NT/1850r.gm_debug <==
> DMA rate for 4096 Byte transfers
> 	bus_read  (send) = 124 MBytes/s
> 	bus_write (recv) = 124 MBytes/s
> 
> ==> proliant-NT/cl1850.gm_debug <==
> DMA rate for 4096 Byte transfers
> 	bus_read  (send) = 128 MBytes/s
> 	bus_write (recv) = 128 MBytes/s
> 
> ==> proliant-NT/snail.gm_debug <==
> DMA rate for 4096 Byte transfers
> 	bus_read  (send) = 128 MBytes/s
> 	bus_write (recv) = 128 MBytes/s
> 
> Here's an ultrasparc, a sun 220R
> ==> sun-220R/i3400.gm_debug <==
> DMA rate for 8192 Byte transfers
> 	bus_read  (send) = 88 MBytes/s
> 	bus_write (recv) = 204 MBytes/s
> 
> A Compaq alpha XP1000, 64bit 33MHz
> ==> xp1000-linux/renoir.myri.com.gm_debug <==
> DMA rate for 8192 Byte transfers
> 	bus_read  (send) = 170 MBytes/s
> 	bus_write (recv) = 240 MBytes/s
> 
> 
> Older 32bit PCI numbers can be found at
> http://www.myri.com/myrinet/performance/DMAperf.html
> 
> If you have Myrinet and GM running on your machine, run
> gm_debug | head -10
> 
> to get the DMA performance measured when the driver loaded the firmare
> on the card.