[Beowulf] Re: newbie's dilemma / firewire? (Hahn)
hahn at physics.mcmaster.ca
Wed Mar 8 09:07:01 PST 2006
> >> effective in some cases. However, the bulk file transfer rates for
> >> very large files via FireWire 400 (1394a) is significantly faster
> >> than 1000baseT and FireWire 800 (1394b) is more than 240% faster
> >> that GigaBit EtherNet.
> > can you show some references for this? it's hard to understand why
> > a 400 Mbps connection would be "sigificantly faster" than a 1000 Mbps
> > connection. or are you talking mainly about shortcomings in some
> > platform's drivers/stack?
> Easiest, simplest reference is this image: http://unibrain.com/
thanks. to me, it looks like the device can manage at most ~24 MB/s,
and both GBE and FireNet can saturate it. that's not too surprising,
since GBE's peak theoretical is 125 MB/s, and if you're getting less
than say 60, you're doing something wrong (85 is reasonable.)
> ... or read the whole report at bottom
> of: http://unibrain.com/products/driverapi/firenet.htm
certainly appears to be saturating because of the disk,
not the interconnect. a P4/1.7 is circa -5 years, and in
those days 23.4 MB/s wasn't atrocious. now, ~50 is about
the least you can expect from a single disk, and that's
not enough to saturate GBE.
> Note that for large files, the FireWire network bulk file transfer
> rates can exceed GigaBit performance.
definitely not shown on those pages...
> The above reports (arguably
> promotional) are for FireWire 1394a (400 Mbits / second). When
> similar tests are run on FireWire 1394b (800 Mbits / second),
> performance can exceed 240% of network bulk file transfer (write
> times) of GigaBit EtherNet.
well, that's what you said before, and I'm still asking for some
indication its true.
I'm mainly curious about the 240% part. oh, wait, are you just
dividing 24 MB/s by 10 MB/s in the table at the bottom? that's
> Why is it comparable to or faster from a processor running at around
> 40% of GigaBit processor speeds? Processor efficiency FireWire has a
> 32-bit "risc" type microprocessor, is peer to peer in hardware /
> firmware and has other lean architecture features. (Small address
> space, data frame large = more efficient data packet over double
> duplex connections.)
that's nice marketing-speak, but content-free, I'm afraid.
> Future: there are prototypes of FireWire 1600 and 3200 operating over
> fiber and using multiple "colors" ... using two fibers (duplex),
> multiplexed by frequency discrimination (4 or 8 channels) using the
> same FireWire 800 (Texas Instruments chips) ... comparable to ~~
> 64000baseT ... and the involved engineers say they can stay ahead of
> Moore's Law beyond several more years.
OK, I get the point. though I have to say I really don't expect optical
WDM firewire to make any difference ;)
what's the latency of a typical 1394 switch? are they all store/forward?
are they all full wirespeed (in internal bandwidth)? what's the max fanout?
I'd personally love to see something better in every way than GBE/10GE,
but just don't see it happening. no, please, don't suggest IB :)
More information about the Beowulf