cheap clusters

Velocet mathboy at
Wed Apr 4 22:04:56 PDT 2001

Thanks to the thoughtful people who replied regarding their configurations.

It just struck me as odd that people were paying 3x as much for DDR nodes. It
isnt actually like that (anymore, SDRAM seem to have climbed a good 10-20% in
the last month or so to begin with and DDR is dropping - perhaps in responce
to SDRAM). Also, I think a lot of people have a lot more overhead PER node
than I am thinking (my setup is a diskless cluster with NFS roots and very
little parallel network traffic). This gives me certain flexibilities others
cant abide by (in fact, the nature of the jobs I am going to be running allow
me to not necessarily need ECC ram or even machines which have a guaranteed
uptime of 6+ months).

It works out that the raw cost of the DDR systems not including
any network, cabling, power, airflow, real estate, cabinet or other
infrastructure is about 2.3 times the price (at least here in the 3rd
world of 'canada' with our prices).

2.3 times is a tall order to beat however, when comparing the pricing
of a 900Mhz Tbird on traditionally SDRAM and a 1.2GHz Tbird on DDR.

With all my infrastructure taken into account that isnt an operational
cost (ie cabinet cost, switch cost, cabling, power supplies, etc), the
ratio drops a bit, to about 2.2. 

The problem is I cant give an accurate comparison. I havent got G98 working
with Atlas yet, and all I have to compare is a few runs on an ASUS A7M266
DDR board with a 1.2GHz athlon with G98 compiled with Atlas.

Even in this disadvantageous comparison, I have the following results:

(the jobs are arbitrary, email me if you want to run them too just
for comparison or something - they're for gaussian 98).

The numbers for the jobs are ratios of 'speed', the inverse of time taken.
(I normalized this all so Tbird 900 = 1.00) - higher numbers are better.

I have some stats for different ram speeds too, but thats not of importance
here. These are all for 133MHz ram, and 266 for the DDR. Recall the DDR
is with G98+atlas whereas the rest are plain ole G98.

		job 1		job 2		job 3 		job 4
		'''''		'''''		'''''		'''''
Duron 700	0.67		0.76		0.69		0.65
Tbird 750	n/a		0.84		0.91		0.89
Tbird 900	1.00		1.00		1.00		1.00
Tbird 1200 DDR	1.26		1.60		2.16		1.94

(incidentally, the jobs are a medium size rhf-321g, 'A' at rhf321g,
'A' at b3ylp and 'A' as a frequency scan.)

The jobs were prepared by the researcher who will be using the cluster,
I dont understand much about the details of gaussian beyond its resource
usage (I have stats for disk and network usage during calculations)
or the types of jobs it runs. The other problem is that the rearcher(s)
themselves are not sure how many of which type of job will get run.

The 3rd job points to the DDR system being defintely equal to the non DDR
system. So it really depends on how many jobs of which type we're doing.
It will either be job type 1 or 3 mainly, apparently, and I cant get
any more exact answers than that right now. As always 'I wish we could
wait a few months to buy' but thats lost time, and some other wierd
price/performance situation will also be happening then with a 'solution'
(price drop in one type of system) right around the corner.

If we end up doing job 1 type jobs, then DDR is a sure loss. If we do job 3,
then DDR is merely equal in price. So we're looking at a price performance,
for my situation, 1.2Ghz TBird/ddr:900Mhz Tbird/SDRAM of job 1,2,3,4 of
0.57, 0.73, 0.98, 0.88.

I remind you these numbers are comparing G98 without atlas to the DDR
board with Atlas. Surely the G98 with atlas on the SDRAM systems is going
to perform at least as well, if not a fair bit better.

It struck me as odd that people are also comparing numbers that
show P4s about 20% faster without mentioning any pricing at all - the P4
is a lot more than merely 20% more expensive. What increase in pricing
is NOT justifiable anymore? 

A low end P4 1.3Ghz system setup the same as the above examples comes out to
almost 3 times more expensive. Is it really more than 3 times faster? Anyone
run G98 yet on a P4 loaded with RAMBUS ram? Are P4's even stable yet?


More information about the Beowulf mailing list