Which Switch?

Robert G. Brown rgb at phy.duke.edu
Sat Aug 25 07:56:43 PDT 2001

On Sat, 18 Aug 2001, Mark Hahn wrote:

> > I could not find any reviews or info about 3com performance.  The review
> > for the netgear I found is located at
> > http://davidcoulson.net/writing/pcp/170/review-netgear.pdf
> this review is a steaming pile of offal;
> it's measurement methods are at best uh, "naive".

Hear, hear.  The writer is an idiot.

I have Netgear's far cheaper 10/100 switches (like the FS108 and FS116)
in both my house and clusters at Duke and they can easily get more than
90 Mbps bandwidth (perhaps 93-95% of theoretical TCP bandwidth) and <
90usec latency when tested with a pair of decent NICs (e.g.  3c905's),
slightly less bw and about twice the latency when tested with a pair of
FS310s.  The only way I can imagine getting only 14 Mbps out of a
>>higher<< end netgear is to use a pair of known-terrible NICs (e.g.
RTL8139s, which tends in my experience to choke on high speed data
streams) at the endpoints or to be testing (by chance) a "broken"
network -- marginally bad cables, poor wiring, or a switch itself that
is overheating or going bad or to e.g. disable autonegotiation so that
the NICs and switch are mismatching duplex or speed.  Or something else

Unfortunately, in addition to using "ftp file copies" to measure
bandwidth (instead of any of the various moderately reliable measurement
tools, e.g. netperf, tools in lmbench, netpipes) the author didn't TELL
us what NICs he was using in the test.  I've (reliably) measured
bandwidths that vary by factors of 3-5 and latencies that vary by
factors of 2-3 for the same switch but different NICs.  If I had to bet,
I'd bet that the author is using $10 rtl8139's to test a thousand dollar
plus switch, at least on the 100BT ports.  I have no idea what he is
doing with the gigabit cards.  He also doesn't say what operating system
he is using or whether the driver he is using is e.g. setting the duplex
and speed correctly during the autonegotiation phase (which can very
easily cause strange/terrible/inexplicable results in a benchmark).

> netgear is a reputable vendor that focuses on
> the low-mid-range market.  I don't know of any real
> performance problems with their products.

I'd go further and endorse their switch products for small to midsize,
not-too-demanding cluster design.  It is possible, of course, that the
1000BT uplink stackable switch he is reviewing is somehow broken in its
design.  And netgear isn't perfect -- I'm still irritated at them for
discontinuing the very acceptable PNIC-based FS310 10/100 NIC (which was
like a very drinkable cheap table wine in a sea of trash on a number of
local vendor counters:-) in exchange for the FS311, which just added to
the trash.  I wouldn't believe it, though, until somebody with sensible
tools (and some sense) tested it and posted the results.

Ignore this article.


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 mailing list