New class of Beowulf clusters ? (Re: two NICs - channel bonding - tradeoff)

[Petr Ladislav Kodym]

On Wed, 2 Jun 1999, Walter B. Ligon III wrote:

>
>Generally I think there are a number of interesting issues in using multiple
>networks.  We have experimented with using one bus and one switched network
>(back when switches were still pricey).

    Hi All,


What I'm really interested in is coupling of fast ethernet with PAPERS
"network".

--------------------------------------------------------------------
PAPERS, Purdue's Adapter for Parallel Execution and Rapid
Synchronization, is custom hardware that allows a cluster of
unmodified PCs and/or workstations to function as a fine-grain
parallel computer capable of MIMD, SIMD, and VLIW execution. The total
time taken to perform a typical barrier synchronization using PAPERS
is about 3 microseconds, including all hardware and software overhead;
this is several orders of magnitude faster than using conventional
networks, and is even faster than most commercial parallel
supercomputers. A wide range of aggregate communication operations are
also supported with comparable efficiency. Despite this performance,
the public-domain PAPERS designs are less expensive than most
conventional networks and are scalable to very large clusters.

ttp://garage.ecn.purdue.edu/~papers/
-------------------------------------------------------------------

Folks in CSC, Helsinki, Finland have measured 312 barriers/sec for 32
processes running on 16 dual PentiumsII 400Mhz connected by switched
fast ethernet (under MPI over TCP/IP). It clearly limits usability of
such a cluster only to (very) coarse grain parallel programs. PAPERS
can handle 330 000 barriers/sec, thats 3 orders of magnitude more. 
They can also transfer short messages with very low latency.
(PAPERS is a small chunk of basic TTL logic circuits which is
connected to the parallel port of PC)
It would push Beowulf clusters to completely new application areas.

Has any of you ever tried this PAPERS thing ?
Can you give any comments ?

** Does someone know, how many barriers can by done over MYRINET or **
** using U-Net ?                                                    **  

Is anyone running code, that is bottlenecked by the need for frequent
synchronization, and therefore possibly interested in trying PAPERS ?



			Petr