[Beowulf] Broadcast - not for HPC - or is it?
Lux, Jim (337C)
james.p.lux at jpl.nasa.gov
Tue Oct 5 14:58:28 PDT 2010
> -----Original Message-----
> From: beowulf-bounces at beowulf.org [mailto:beowulf-bounces at beowulf.org] On Behalf Of Greg Lindahl
> Sent: Tuesday, October 05, 2010 1:35 PM
> To: beowulf at beowulf.org
> Subject: Re: [Beowulf] Broadcast - not for HPC - or is it?
> On Fri, Sep 24, 2010 at 08:21:55PM +1000, Matt Hurd wrote:
> > This was not designed for HPC but for low-latency trading as it beats
> > a switch in terms of speed. Primarily focused on low-latency
> > distribution of market data to multiple users as the port to port
> > latency is in the range of 5-7 nanoseconds as it is pretty passive
> > device with optical foo at the core. No rocket science here, just
> > convenient opto-electrical foo.
> If you go read up about the Blue Gene series of machines' networks,
> one of them is a "Eureka" network for global broadcasts. It's only a
> minor aspect of most scientific computations, though. There was even a
> very low cost, low-latency broadcast network out of Purdue called
> PAPERS that used the unused parallel port that used to be available in
> most servers. It was pretty amazing what they could do for so little
> $$, but I don't think they found that many applications.
PAPERS was pretty neat, but these days, there are fewer motherboards with a parallel port, and even fewer with a "well behaved" parallel port suitable for PAPERing..
You'd also have a tough time getting latencies down in the sub microsecond range, since the parallel port is fundamentally intended to talk to a "Centronics" printer interface, with 1 microsecond setup, 5 microsecond strobe, and 1 microsecond hold time, as I recall. (Plenty fast running to that line printer at 400 characters/second, eh?) The EPP and/or ECP found in more modern equipment runs at maybe a megatransfer/second. You're still limited by the equivalent of LS244 and LS374 kinds of speeds and loads.
More information about the Beowulf