Re: MP on P5 vs P6

>[David Jeske <jeske@home.chat.net> writes:]
>
>>What's your feeling and experience on P5 SMP vs P6 SMP? It's very
>>inexpensive to get the required equipment to build a P5 SMP system ($190
>>for a Tyan Tomcat 4 dual P5 board, $150 for a P5/166MMX CPU). While a P6
>>is somewhat more costly ($290 for a SuperMicro dual P6 board, $275 for a
>>P6/180), I would expect it's "SMP" related performance to be significantly
>>higher, due both to the better processor bus protocol, and to the dual,
>>and private L2 caches.

I've seen PPRO-150s around here for $215 each, and can almost always
be throttled to -166. The trouble with the -180s is that you must
run a memory bus speed of 60Mhz (180/3) and thus a PCI busspeed of 30Mhz.
If OTOH you can run -166 or -200 (expensive) you can strap the memory
bus speed to 66Mhz (166/5) and PCI speed of 33Mhz. From what the
*BSD crew have been saying it seems the 66Mhz multiple processors
(133, 166, 200 aka 198) tend to be a better bet if you're running
stuff that bounces out of the cache (e.g X11 :) ) or if you want
the most out of your I/O. I think the speed bump I saw going from
-150 to -166 on my PPROs was about 10-15% but that's purely subjective.
Plan9 certainly runs faster :)

Unfortunately PPRO-150s are getting increasingly hard to find.

I've had zero trouble with the Tyan S1668 ATX 2xPPRO board running
everything from FreeBSD-3.0-SNAP (SMP) to WinNT4. The SuperMicro
boards are supposedly pretty stable too (except yours, Andy? :< )

>>Of course these are all the bargan basement boards and CPUs, and I would
>>guess there would be significant gains to using 512k L2 instead of 256k L2
>>P6 chips.
>Yes, but a coherent L2 cache at 256k is still usable.

I've found some vendors who didn't know how to tell externally
between a PPRO200-256k and PPRO200-512k. Needless to say they'll never
be seeing my business again... one has to be careful about these guys.
I'm sure there are numbers on the chip somewhere, but I'm still not
absolutely sure. AFAIK the 150s were only built with -256k L1s.

[xx]
>My recommendation would be to go PPro for sure. If you want to economize,
>getting the 256k cache configuration is a better compromise than going to
>P5's. Shared L2 cache with write-through L1 is going to make it hard to
>scale to any extent at all.

I think 'disappointing' was the word used when the FreeBSD crowed surveyed
P5-SMP performance :) Their caches just aren't big enough I guess.

> Andy

Regards,

Eric Dorman
edorman@tanya.ucsd.edu
Received on Wed Sep 10 07:16:09 1997