sorry ... not cut through
Robert G. Brown
rgb at phy.duke.edu
Tue May 14 09:18:56 PDT 2002
On Tue, 14 May 2002, Donald B. Kinghorn wrote:
> I just tried to get some more information about the Netgear switches and
> according to there web site the 524 is "store and forward" and not "cut
> through" like I reported. The FS508 was a " cut through" design but I see
> it's not even listed as a product anymore on the Netgear web site.
> When I use inexpensive switches I generally get one with at least twice the
> number of ports than what I actually use assuming that the switching fabric
> will not handle more than full load on half the ports at a time. The FS524
> Netgear has a 1.2 gigabit/sec switching fabric and thus would not be a good
> choice for someone that needs to use all 24 ports ...
Although they do have a slightly more expensive, 24 port rackmount
stackable (2 uplink) switch with 4 Gbps bandwidth. Depending on how the
bandwidth is available to the 100BT ports, this has enough bisection
bandwidth to run 12 full duplex pairs both ways.
Even the 1.2 Gbps switch has enough for 12 pairs being used in one
direction at a time, which likely suffices for many purposes. I
certainly see no saturation on a 16 port switch, but don't have a 24
port switch to test.
The 16 port switch typically delivers 84-90+ Mbps TCP stream (for large
messages, of course) as measured by netpipe. The total TCP latency
appears to be between 50 and 80 microseconds (also from netpipe, for
small messages). The choice of NIC appears to have more impact on
performance than the switch load -- I see signficantly lower throughput
between nodes equipped with cheap NICs (PNICs) than with e.g. 3c905s or
eepros, and although I no longer have RTLs in anything to test, they
were truly terrible performer (which is WHY I have a drawer full of
RTL's that I give away to people I really don't like that much --
Windows users, for example:-).
These numbers aren't spectacular, but they aren't terrible. They are
(as Greg said offline:-) "boring". This is fine. Many, many cluster
applications call for a boring, cheap switch that you just plug in (to)
and it works. The netgear(s) seem to fit the bill, although I'm sure
that linksys and dlink and others have offerings that do as well.
Regarding cutthrough vs store-and-forward -- some years ago this issue
was discussed on the list and I grew curious as to who made cutthrough
switches, which seemed likely to do better on bandwidth if perhaps worse
on latency. After spending a half day with a browser and search engine,
I concluded that basically "all" the cheap switches and a remarkable
number of the more expensive ones were store and forward. This may have
changed as hardware has evolved, but then, it may not have.
Robert G. Brown http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567 Fax: 919-660-2525 email:rgb at phy.duke.edu
More information about the Beowulf