Re: [myrinet] RE: Re: RE: PCI bus speed

On Thu, Aug 30, 2001 at 10:49:26PM -0700, Bob Felderman wrote:
> On Thu, 30 Aug 2001, Tenhave, Tim wrote:
> >performance may not be affected depending on your app's compute versus
> >communication needs.
> >
> >As for Sparc based systems from Sun.  They do not have 64/66MHz systems
> >until you reach into the Ultra 60 workstation class or higher (As far as I
> >know).
> >
> >Note: If this comment is not correct, PLEASE correct me for the benefit of
> >this list.  I would not and will not intentionally make a false statement
> >about a fellow system vendor's product line.
> >
> 
> Since I don't work for any system vendors, I don't have to be so
> politically correct.
> 
> Tim is (of course) completely correct in stating that interconnect
> benchmark performance may not have direct correlation to the
> performance of your application. I believe this thread began with
> someone looking to "double performance" by adding another
> Myrinet card. Clearly there should be some measurements
> of the system/application that points to network bandwidth limitations
> before trying to add another card.
> 
> As for system bashing, I'm happy to join in... ;-)
> Sun may make 64bit/66mhz PCI systems, but Sun forgot how
> to build good I/O systems after the sparc-2. Actually, that's
> a little unfair since they appear to be optimizing for
> server performance rather than peak performance for
> a single application. Compaq's (well DEC's) 64bit/32MHz
> PCI was available MANY years before any other systems
> that were affordable and their performance was quite
> respectable. I'll have to check my measurements, but
> DEC's 64bit/33MHz peak bandwidth meets or exceeds
> what I've seen on Sun's 64bit/66MHz and it certainly
> beats the numbers you can get from an x86 using the
> Intel 840 chipset on 64bit/66MHz PCI.  The 840  has
> lousy PCI read performance. The translates to lousy
> network send  bandwidth.
> 

Sorry, but this is a non sense for me. If you could measure 
better performance on a pci 64/33 than on a 64/66, well 
you are a wizard. 

I see that bw on our (old - around '97) alphas is:

1. non symmetric: 'write' bw is one half of the 'read' one
2. max bw is almost one half of the theoretical bw of a Myrinet 2000
card.

(cfr. Low bandwidth thread for performance values)

Anyway my comparison is not among Dec and Sun. 
I'm interested to bus performances. So if someone could 
prove that Banana Republic Inc. systems with a Linux supported processor
can performance better with GM
than what I see on alphas, well "Great! I'll buy it!"

> If any of the DEC engineers are still working for Compaq,
> then I would assume that that Compaq's 64/66 and
> (maybe sometime PCI-x) will be good.
> 
> Bob
> 
> 

-- 
Francesco P. Lovergine