Mathematics of gigabit question

Jared Hodge jared_hodge at iat.utexas.edu
Fri Dec 7 09:12:49 PST 2001


Mark,
	These are actually very good questions that a lot of people have, which
is why I decided to CC the beowulf mailing list.  Perhaps someone else
on the list could do better, since I'm not an expert on PCI or GigE but
I'll give it my best shot at answering your questions.

First, I'm afraid Intel over idealizes their PCI numbers just a bit
(plus I think they are going with a strange definition for MB).
OK, 1 MB is 1024*1024 bytes = 1,048,576 bytes.  For some reason
manufacturers (especially hard drive builders) tend to go with an even
1,000,000 bytes and pretend that it's 1 MB (actually I know the reason,
and it's not because the math is easier, its because it makes their
products look bigger).  Now, where does Intel get their numbers?  Here's
what I think they did:

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).  

Similarly, you get the 256 MB/sec (really 251.7 MB/sec), and 528MB/sec
(really 503.5 MB/sec) for 64/33 and 64/66 PCI respectively.  Ok, that's
their bad math, now the over idealized part is that they may not be
telling you that this is shared between all PCI slots on the same bus
and that it is for both directions (total bandwidth for the PCI **BUS**
-meaning shared).  Also, no matter what specification is given, you'll
never get full connection speed over any link because of various
overhead costs.  Measuring actual communication speed (using a tool I
got from Myricom and motherboards we actually have) I get:

32/33 = 128 MB/sec (really 122 MB/sec)
64/32 = 250 MB/sec (really 238 MB/sec)
64/66 = 512 MB/sec (really 488 MB/sec)

Note, just multiply 1,000,000 byte megabytes by 0.95367431640625 to get
actual.

Ok, now before I launch into GigE cards there is one caveat.  When you
are going from a PCI connection to any network connection, you are
talking about a totally different type of communication protocol.  I
haven't studied the intricacies of PCI protocol, but knowing all of the
overhead TCP/IP has, the conversion takes time.  This means buffering is
required while processing occurs on the NIC and extra processing is
required on the system processor.  My point is that this is not a
one-to-one conversion so we are glossing over quite a few unknowns. 
Maybe someone else on the mailing list could give you a few more
details.  My guess is that the PCI bus has less overhead than the NIC,
but to be honest I don't know for sure.

I believe most ethernet devices have the ability to work in either full
duplex mode or half duplex modes (I think these terms are a little
weird, especially half duplex.  Seems like it should be just duplex and
not duplex at all, but that wouldn't sell NICs).  That's 1000 Mbps each
direction or 125 MB/sec (really 119 MB/sec) one way, 250 MB/sec (238
MB/sec) in full duplex mode.  So to try to compare apples to apples, for
a half duplex link you've got (I'll use real MB, since I refuse to
conform to marketing ploys):
32/33 PCI = 122 MB/sec vs. half duplex GigE = 119 MB/sec.

Seems like it should work, right.  Well the problem is that aside from
the unknown overhead costs that I mentioned above (which could already
mean the GigE NIC is getting starved a little), we have to figure out
where that data is coming from.  If you want to sustain the full-speed
link for any length of time with real data, you've got to get lots of
data from somewhere which means it's probably not all in physical RAM
(it very well could be, but we don't want to depend on this when
designing a system).  That means the hard drive is probably working
some.  Well, with most Intel PC chipset designs, this goes through the
south bridge (or whatever they are calling it now).  You would have to
look at some board-specific diagrams of your motherboard to know for
sure, but this often means that any traffic from the Hard drive goes
through the PCI bus since the PCI bus connects the north and south
bridges.  On newer motherboards this isn't a problem since there is a
separate connection from the north bridge to the south bridge (I think
that AMD calls it Hypertransport, I don't remember what Intel calls
it).  Again, I know north bridge and south bridge aren't the latest
terms, but it's gotten to the point where even the names for motherboard
components are just marketing.  Anyway, you're bound to be doing
something over the PCI besides just communicating with the NIC, so I
imagine that the NIC isn't going to be fully fed, but I don't have any
hard numbers to give you for this.

Obviously, a NIC operating at full speed on full duplex on a 32/33 PCI
doesn't have much of a chance of staying completely fed.  I imagine
you'd do fine with 64/32 since the chances of actually need full speed
on full duplex for any length of time is very slim and the PCI could do
a pretty good job of feeding it anyway.  If you're going with a really
high end NIC though, it makes sense to keep it fed as well as possible,
which may even mean 64/66 PCI.  I imagine a lot of GigE NICs that are
half duplex are only 32/33 and that are full duplex are only 64/32.  Why
make a more expensive NIC when 99% of your market wouldn't know the
difference and you can already say it's GigE?
	Did I mention the marketing ploys involved in all this?  Actually I
think it's funny how dumb the big manufacturers think the public is. 
Whether when designing NICs (not dumb in this case, just maybe
uninformed), or (very dumb in this case) showing us aliens that are
flying in space that are mystified by the power of a little chip that
they can drop in strange goo (wish I had some, might be useful) and then
do all sorts of wonderful things with, like get stereo sound (wow my
speakers just got better) and edit pictures (wow my software just got
better).  Those earthlings down there sure have advanced technology... 
At least they don't have people flying all over the place for no reason.
	

> "Mills, Mark" wrote:
> 
> I read some of your posting on "64-bit PCI - 66 Mhz vs 33 Mhz
> networking performance?" at http://www.beowulf.org
> If you could answer clear up 2 question that has been bothering me I
> would appreciate it.
> 
> 1.      I have often heard that a 32 bits/33 Mhz PCI bus cannot fully
> utilize a gigabit ethernet card.  But if a 32 bits/33 Mhz PCI bus can
> handle an aggregated ideal peak at 132 MB/s.  (32bits x 33MHz =
> 1056bits or 132MBytes/s) then why not?  A gigabit card only passes
> 125MB/second (1000bits\8bits= 125MB's) right? Or do they mean it can
> only work in half duplex mode and not full duplex for a total of
> 264MB/s
> 
> 2.      If a 32 bits/33 Mhz PCI bus can handle 132 MB/s.  (32bits x
> 33MHz = 1056bits or 132MBytes/s) does that mean that when the NIC is
> running in full duplex mode that the data being sent is running at
> 66MB/sec and the simultaneous data being received is at 66MB/sec for a
> total transfer of rate of 132MB per second?
> 
> Thanks for any help you can give.
> Mark Mills
> 
> Voice 281-370-3861
> Fax     281-370-3801
> Email  Mark.Mills at Desktop-Assistance.com
> 
> Intel's site gives this info on Gigabit ethernet and today's PCI
> slots.
> You need to keep in mind the (theor.) peak of the PCI bus :
> 32 bits/33 Mhz : aggregated ideal peak at 132 MB/s
> 64 bits/33 Mhz : aggregated ideal peak at 264 MB/s
> 64 bits/66 Mhz : aggregated ideal peak at 528 MB/s"
> Found at
> http://support.intel.com/support/network/adapter/1000/sb/1010453072955946.htm

-- 
Jared Hodge
Institute for Advanced Technology
The University of Texas at Austin
3925 W. Braker Lane, Suite 400
Austin, Texas 78759

Phone: 512-232-4460
Fax: 512-471-9096
Email: Jared_Hodge at iat.utexas.edu



More information about the Beowulf mailing list