Mathematics of gigabit question
lindahl at conservativecomputer.com
Fri Dec 7 11:55:34 PST 2001
On Fri, Dec 07, 2001 at 11:12:49AM -0600, Jared Hodge wrote:
> 33,000,000 cycles/sec (that's 33 million cycles per second or 33 Mhz) *
> 4 bytes/cycle (32 bits = 4 bytes) = 132,000,000 bytes/sec (132 small
> MB/sec, or 125.8 real MB/sec).
Nope. 33 Mhz is really 33.333333 MHz, and so the actual sustained
bandwidth is 133 real MByte/sec. If you look on my Myrinet experiences
page, you'll find the actual sustained performance of PCI busses
talking to Myrinet cards is (of course) less than ideal, but some
motherboards do quite well. The best example of 33mhz/32bits is 128
real MBytes/sec, and at 64bits/66mhz is 455 "real" MB/s.
"small" megabytes are only used for disks by the marketing (and now
technical) people. They mostly haven't invaded the rest of the
industry, so I've yet to see a technical person use "small" megabytes
for anything but disks. But you need to know the actual clock rates;
everyone rounds them.
It's likely that gigE cards don't have a significantly better
implementation of the card/PCI interface. But of course processing the
TCP stack is extremely expensive, so that's the main limit.
More information about the Beowulf