Petabits/sec, and the like
James.P.Lux at jpl.nasa.gov
Wed Nov 5 12:40:09 EST 2003
All of the enjoyable chat about achieving stupendous data rates with disk
drives in trucks is quite interesting. By the way, I don't know why you
insist on having the drives mounted in racks..why not just leave them in
their original shipping containers. There's also the concept of how many
bits are being moved in, say, a container load of Britney Spears DVDs.
(leaving aside questions about redundancy, information entropy, and whether
there is any information content in Britney Spears to begin with)
But, on to a more practical aspect. It seems that a mere bits per second
number isn't useful, because it doesn't embody some practically important
things, like latency or transport time, both of which can be
significant. This is of particular concern to me, because I'm used to
having to deal with networks where the round trip light time is significant.
So, I propose that an interesting single metric might be to scale the bit
rate by the latency with which the bits appear at the other end of the
pipe. As illustrious an early high performance computing as Seymour Cray
recognized that this could be significant when you're looking at pumping
lots of bits real fast.
And, there's a handy yardstick to measure by (issues of quantum
entanglement and photon twinning aside), in vacuo speed of light.
For example.... old style 10Mbps thinnet ethernet used solid dielectric
coax, which had a propagation velocity of about 0.66 c. twisted pair is
probably around 0.75, fiber optics are a bit tricky, depending on the mode
of propagation, but probably around 0.85. The pickup truck full of disks
is about 1E-7. The units of the new measure would be, what, (bits per
second)*(meters per second) or bit meters per second squared. I'd normalize
by c, to make the units more useful..I'd modestly propose calling the new
unit the Lux, but it's already been used, so perhaps we should recognize
rgb's contributions by calling it the "Brown" 10Mbps over thinnet would
then be 6MegaBrowns. 100mbps over twisted pair would be 70MegaBrowns. The
1 Pb/s truckload of disks would be 100MegaBrowns.
This is clearly the "raw pipe speed" too... not taking into account the
headers and any coding that's going on. The disk drive pipe hides all the
coding and sector headers, so the measurement is a real data transfer
throughput. The Ethernet scheme on the other hand, is just the signalling
rate, and there is some significant non-zero overhead.
One might also ask whether physical size of the system being communicated
within should be factored in (say, when talking about bisection
bandwidth). Clearly, a cluster with a physical dimension of 100meters is
going to be slower than one with a physical dimension of 1 meter, all other
things (processor speed, comm speed, etc.) being equal.
One has to also consider the bandwidth of the entrance and exit to the
pipe... merely having the capability to transport Tb of disk drives rapidly
doesn't mean that you can put data onto those disks at a Pb/s and get it
off at the other end of the shipping channel. This is where those "use
free air as a communication medium" schemes get into trouble. Sure, the
optical bandwidth of air (or optical fiber) is pretty darn wide (on the
order of 0.5 PetaHertz (a unit I never thought I'd ever use) for just the
visible spectrum) but the modulation and demodulation might prove to be a
There's also the issue of real computing efficiency.. speed is not
everything in some applications... some applications might optimize for
calculations per Dollar/Euro or calculations/Joule. Coming up with a
metric for the calculation is a bit tricky. The calculations could be
viewed as extracting information bits from a redundant data set (a
coding/decoding process), or as creating new information (although, hmmm...
this gets a bit metaphysical)
I leave the selection of appropriate units and names to the community.
James Lux, P.E.
Spacecraft Telecommunications Section
Jet Propulsion Laboratory, Mail Stop 161-213
4800 Oak Grove Drive
Pasadena CA 91109
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
More information about the Beowulf