[Beowulf] Gigabit switch recommendations

Tim Mattox tmattox at gmail.com
Thu Mar 30 06:04:23 PST 2006

Hello fellow Beowulfers,
IMNSHO, Marketing Departments should be jailed... or at least not
allowed to do the "technical specs" pages for networking products.
And at a minimum, they need to be flogged with a wet noodle... :-)

The following technical buzzwords used[1] to mean something when
they were included on the "technical specs"  (now turned into
"glossy brochures") for a network switch:
  wire speed
  full backplane bandwidth
  X Gbps backplane bandwidth (replace X with some random number)
  X Gbps aggregate bandwidth

These days, the above are pretty much the buzwords/buzphrases that
are on every commodity switch even if they don't apply. (Along with
various dubious management "features".)  See this switch which I looked
at earlier this week... It's technical specs clearly have an error relative
to the marketspeak on the same page:
At the very top it says:
"Non-blocking bandwidth"

Yet near the bottom of the same page it says:
"Bandwidth: 40 Gbps"

It's a 48-port Gigabit Ethernet (on every port) switch...  for it to
truly be non-blocking, it would have to have at least 96 Gbps bandwidth.
(Or 48 Gbps if you are not counting in both directions...  but AFAIK,
you do need to count for bidirectional... and thus 96 Gbps of
internal bandwidth would be required to be non-blocking on 48 GigE
ports.)  Now, this could just be a simple typo, where the tech specs
from a smaller 16 or 24 (20?) port switch were copied...  but that still
wouldn't be right, and well, just copying tech specs from one device
for use in another brochure seems very wrong to me.  Why include the
tech specs if they are not for that device...

But if they had said "96 Gbps", I still wouldn't really believe them, since
I've seen too many where it was obvious the marketers just multiplied
the port count by 2 Gbps, which had no correlation to the performance
of the internals of the switch.

Does anyone know of a magazine or organization that regularly
benchmarks and tests these performance claims for the
*commodity* networking industry?

I know about the Tolly Group and http://www.networkworld.com/
which periodically are commissioned to benchmark the latest
& greatest switches, etc. from the name brand vendors (latest I
see is a test of a 10 Gbps/port 24-port Force10 switch, which was
of course very impressive...  and way out of the price-range for
the clusters I'm typically asked to help design).

What I've been looking for is benchmarks of the switches
that are under $1000. Anyone know of any that have been done?

I wish buying network equipment was like buying orange juice, where
it really didn't matter much which brand you got, and you could
read the carton to look for things like fortified with Calcium, or whatnot...
and then look at the pricetag for $/ounce and make your choice.

Anyway, off my soapbox, and back to dissertation writing...  I'm
almost done... but not done yet...

Footnote [1]: Maybe those buzzwords never really meant anything,
and it is just me being foolish thinking they did at some point have
meaning...  our original 32-port Fast Ethernet switches on KLAT2 back
in 2000 had all the right buzzwords... yet were clearly not capable of
their claimed performance (I checked the true tech specs of the chips
they were built with, once we had purchased them, and, well, after they
started dying.

But that was years ago, and those switch carcasses (um, evidence?)
are long gone...  Hmm, didn't I say I was getting off my soapbox...  ;-)

FYI - here are some links to a variety of 48-port commodity GigE switches,
that may be worth looking at (but as another poster indicated, these
might actually all be the same switch built by an OEM and just rebadged.
One can only really tell by opening one up and looking at the PCB and chips
that are inside.):


D-Link DGS-1248T

SMC8648T TigerSwitch

TRENDnet's TEG-448WS

(This is the one with the funky tech spec mismatch I talked about above.)

On 3/30/06, Bruce Allen <ballen at gravity.phys.uwm.edu> wrote:
> > On Tue, Mar 28, 2006 at 10:01:03PM -0600, Bruce Allen wrote:
> >> After discussions with some of their technical people, the conclusion was
> >> that the switch had enough buffer memory to run wire speed with MTU=1500
> >> but not with MTU=9000.
> >>
> >> Our testing results are here:
> >> http://www.lsc-group.phys.uwm.edu/~parmor/Network-performance/netperf.html
> >
> > But according to that graph, the Dell 2716 switch gave performance
> > substantially below wire speed for ALL MTU sizes, including 1500.  So
> > what's up with Dell's "wire speed but only at MTU=1500" story?
> You'll have to ask Dell! I agree that their explanation does not make
> sense, for exactly the reason that you gave.  We didn't see wire speed
> performance even at MTU=1500.
> Cheers,
>         Bruce
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

Tim Mattox - tmattox at gmail.com
    I'm a bright... http://www.the-brights.net/

More information about the Beowulf mailing list