Mathematics of gigabit question

Velocet math at
Sat Dec 8 10:56:24 PST 2001

On Fri, Dec 07, 2001 at 05:38:43PM -0500, Patrick Geoffray's all...
> Jared,
> Your post was a good guideline.
> Two more things to keep in mind about PCI buses:
> * PCI is a bus, that means you want to reduce the 
> number of devices to the minimum if you want maximum 
> performance. Each time you shared a PCI with something, 
> you will of course give away some time on the PCI, but 
> there will be a lot of PCI cycles wasted in arbitration.
> Also, the PCI settings depend of the weakest devices on 
> the bus: if you plug a SCSI controller supporting 64 bits 
> and 33 MHz, your bus will be set at 33 MHz and everybody 
> else will run at 33 MHz, even if the bus and another device 
> can run at 66 MHz.

This is what I was thining of when I was reading up on those "PC's on a card"
things on Slashdot in the last couple months.  One of them was talking about
communicating with the host machine via a private virtual network over the PCI
bus. I was wondering what kind of latency and bandwidth you'd get on
someting like that, with 5 or 6 pci-mounted PCs trying to all chat
with one another.

> * There is some overhead in the PCI protocol, to get the right 
> to speak on the bus or to tell the PCI chipset where the 
> data you are goign to write will go for example.
> This overhead is defined in PCI cycles. A 66 MHz bus will have 
> 2 times more cycles than a 33 MHz bus. The 64 bits/32 bits part 
> is the size of the data transfered per cycle in burst mode,
> 64/33 is not the same things than 32/66. In a 66 MHz PCI bus, 
> the overhead is smaller. 
> PCI-X is a 64/100 PCI bus, plus some great features like 
> interlaced transaction.
> Motherboard vendors are not very interested in PCI performance, 
> because the market of people able to max out a cheesy 32/33 PCI 
> is tiny. Video was a killer application, but it was moved to a 
> separated bus (AGP), mainly for marketing reason I think.

What kind latency and bandwidth can be achieved over IP over SCSI (apologies
if that's been talked about here recently).


> Patrick
> ----------------------------------------------------------
> |   Patrick Geoffray, Ph.D.      patrick at 
> |   Myricom, Inc.      
> |   Cell:  865-389-8852          685 Emory Valley Rd (B)
> |   Phone: 865-425-0978          Oak Ridge, TN 37830
> ----------------------------------------------------------
> _______________________________________________
> Beowulf mailing list, Beowulf at
> To change your subscription (digest mode or unsubscribe) visit

Ken Chase, math at  *  Velocet Communications Inc.  *  Toronto, CANADA 

More information about the Beowulf mailing list