Linux Cluster attains 125.2 GFLOPS

akey writes "CPlant98, a Linux cluster composed of 400 Digital Personal Workstation 500as, achieved 125.2 GFLOPS, which would place it at #53 on the Top 500 list. And this was only a 350-node run... " I'm hearing rumors of 1000+ Linux Clusters. I'm itchin' for it to come out of the closet so we can see some real benchmarks.

Comments

Cluster Scaling

[_Quinn]

How large can a cluster be?

Short answer: it depends.

Long answer: it depends on the applications and the usage patterns.

(I'm assuming we're talking about practical limits here, not theoretical ones -- the theoretical limit is probably the address space of a cluster's message-passing interface (i.e. 4 billion nodes).)

Some applications -- the so-called "embarassingly parellel" ones -- will scale with nearly no deviation from linear to any number of nodes, because they do loosely-coupled problems. (Which means the result of one part of the parellel computation does not depend on a result from some other parellel computation. The mandlebrot set is a good example of this.)

In general, the more tightly-coupled the problem is, the harder it is to scale, as the amount of data that has to be exchanged pushes the limit of the interconnects. A 32-node cluster constructed on a hub will be faster for loosely-coupled programs than a 24-node cluster on a switch, which could beat the 32-node cluster on a tightly-coupled problem because of communications overhead in the 32-node cluster.

Usage patterns also determine the maximum useful size. If you're at a large lab like Sandia, you can reasonably expect a large number of jobs to be running concurrently, which essentially parellelizes the cluster -- running 6 tightly-coupled programs, each on their own hypercube interconnect, will complete faster than running the six in series, each with the whole cluster.

-_Quinn