AMBER Archive (1999)

Subject: Re: PC clusters running AMBER

From: dek_at_cgl.ucsf.EDU
Date: Tue Oct 19 1999 - 10:26:05 CDT

Xavier Deupi writes:
>Hi all,
>
>We are planning to build a beowulf-type cluster to run AMBER. We wonder if
>there's people doing it, if there's some benchmarking,...

There are some PC clusters running AMBER. For one, there's the LOBOS cluster
at the NIH (see http://www.lobos.nih.gov).

I have been trying several technologies for building very a low-cost
AMBER cluster using PCs running Linux, but I have found that scaling
beyond just a few nodes is challenging, due basically to Amdahl's law:
a parallel algorithm can only go as fast as its serial sections proceed,
and with the low bandwidth and high latency of PC bus/network technology,
you rarely get a speedup of better than 70% for a second node.

One alternative, although expensive, is to go with a faster networking tecnology
(mine is just 100BT, and I only get about 6MB/s data transfer for large messages).
Myrinet and M-VIA are the current competitors. Both feature low latency (for
small messages), high bandwidth (for large messages), and user-space MPI
implementations which can avoid the problems associated with the Linux TCP stack
(for one, the TCP stack until recently had errors in the packet fragment
reassembly, where the Nagle algorithm was being applied to the final packet,
leading to HUGE latencies!!)

"Benchmarks":

For my system, a 24 residue DNA duplex in a box of water ~ 50X50X50 A,
I get about a 70% speedup for using the second processor (this is on a
dual Pentium II 400) and 100% speedup using four processors (2 Pentium
II 400 connected with 100BT). This is not exactly "great" scaling.

Dave
-----------------------------------------------------------------------------
Email: dek_at_cgl.ucsf.edu David Konerding WWW: http://picasso.ucsf.edu/~dek
-----------------------------------------------------------------------------