(osum tucnaku)

Comparisons Considered During Design of Linux Cluster MAGI

Is it better than (network of) workstations?

First of all we considered what we could get from traditional workstation vendors like HP, SUN, SGI and IBM. We wanted to compare performance/price ratio for workstations and PC cluster. I must admit that these numbers are very hard (if even possible) to calculate. It is not clear what to put in for performance - there are many different benchmarks - and it is even less clear what to put in for price - it might depend on quite unique discounts. Furthermore, both quantities are very quickly changing in time. Due to this situation, every vendor claims the ratio is best for his solution and I am doing to claim it for Linux PC cluster as well :)

DISCLAIMER: all numbers in this paragraph are just results of my personal computitions made on randomly gathered data during my decision making process. If you are workstation vendor, please note that I might still like you product, I just consider it to be bad solution for some of our particular needs.

The only chance to get some meaningful performance numbers is to consider what exactly are you going to do with your system and concentrate on parameters which are important for it. We knew from experience that for our typical demanding task we need quite a lot of disk space, where size is more important than speed (it is much better to have whole database on slow disk than half of it on quick one). We need as much CPU power as we can get, but we do not need extremely quick inter-CPU communication. We know our big tasks can live in some 64 MB of real memory without trashing. We furthure know that floating point and integer performance are of roughly the same importance for us.

With this in mind, we evaluated SPECint95 and SPECfp95 per US $1000000 (spring 1997) and MB of RAM and GB of diskspace per $1000.

system SPECint95/M$ SPECfp95/M$
HP VISUALIZE B160L,32MB,2GB,17",Sep30/96 * 457 445
HP VISUALIZE B132L,32MB,2GB,17",Sep30/96 * 544 572
HP VISUALIZE J280,64MB,2GB,Sep30/96 306 501
HP J210XC 1proc.,Jan20/97 524 898
HP J210XC 2proc.,Jan20/97 674 1154
Sun Ultra 1/140,Jan20/97 522 ?
SG O2 R5000 SC,64MB,4GB,20"c 562 673
SG O2 R5000 SC,32MB,1GB,17" * 1125 1265
whole MAGI as described here,May/97 3170 2442

* Systems with 32MB are quite unusable for us and are here just for comparison.

Additional RAM was 5.6 times cheaper for PCs than for HP machines at the fall of 1996 (however HP uses quicker memories, this compares just size).

To be realistic, we also considered which system we COULD buy for the same money instead of MAGI. Among the best, we could buy for example HP J210XC, 64MB RAM, 2GB HD. The following table summarizes how many times more individual 'qualities' we got in MAGI:

5.9x more 2.7x more 16x more 39x more

The HP system was chosen to try to beat MAGI's weakest point: floating point performance. This created quite big difference at the other end - HD size.

Linux PC cluster solution offers at least three times more for the same money.


Network interconnect

Lantimes about SuperStack II Switch 3000

Comparisons with supercomputers

This is interesting mainly as comparison of architecture type - in principal it is similar.

It is worth noting that today's fastest supercomputer in the world Intel's ASCI Red also uses Pentium Pro (here is more technical info).

Here is current list of top 500 supercomputers in the world. You can also see just pictures of theese machines.

Back to main MAGI page
This page was created and this WWW server is maintained by Vaclav Hanzl