Author: Vincent Diepeveen
Date: 19:44:36 09/11/01
Go up one level in this thread
On September 10, 2001 at 22:05:54, Robert Hyatt wrote:
>On September 10, 2001 at 17:39:40, K. Burcham wrote:
>
>>
>>
>>could you comment about this monster. is a setup like this capable
>> of running a single chess program like we know it today.
>>
>>what would you guess this thing could hit in kns?
>>
>>it seems at these greater depths a human would have a worthy
>>opponent.
>>
>>could this hit about 25,000 kns? Robert
>>
>>wonder what depth this would be in normal tournament time control?
>>
>>also did i read that page correctly? are they offering 64 individual
>>1200 amd processors for only $70,000. slate if you read this
>>i want first bid on the new dual 1400 you just purchased.
>>
>>
>>
>> http://www.microway.com/products/clusters/dualathlon.html
>>
>>kburcham
>
>
>The link says it all. "clusters". In a short answer mode, "no, chess programs
>won't be able to use this platform." I have plans for clusters, but nothing
>functional yet...
>
>I'm not aware of anybody that is really doing clusters today.
I had plans, but i stopped when i realized that
a) i could only get system time from you if i was lucky
b) no one else would be able to give me system time all so called
people who build clusters, own clusters, are admins on clusters or
whatever and who i asked for system time there in the end
i heart nothing from them to develop at
c) it's overwhelming difficult to get a positive speedup at a cluster
considering nowadays search depths. In past depths above 8 were more
than what a normal person got and also no one asked you with what kind
of program or search method you got the depth, but times are changed.
d) usual nodes from a cluster, with exception of this cluster are dead slow,
like 300 or 450Mhz machines. I have a dual 1.2Ghz AMD now myself and
it's very fast.
e) the speed calculations are nearly always disadvantaging clusters.
for example one of the fastest clusters i know where a speedup might
be able to get is your machine: 8 x 4 x 550Mhz (katmai core).
Now first let's use the 30% reduction or 0.7 multiplication for
compensation that a K7 MP is like 30% faster than a katmai core.
0.7 x 8 x nodes
Now speedup one would get on a cluster is most likely not
even near 50%. I would estimate it at 33%:
0.7 x (8 x .33) x node
Now the speedup at a single node is already only like 3.1 or something,
for me it's a bit more depending upon what extensions i use. Right now
i use all kind of stupid extensions that could find some weird tactics
like we have in testsets, so let's go for 3.1 here.
0.7 x 8/3 x 3.1 x 0.55Ghz = 3.18Ghz
So a cluster which on paper delivers 32 x 0.55 = 17.6Ghz i already
estimate at getting 3.18Ghz for me. So if i wait a month and equip
a dual AMD with 1.6ghz processors i will be already way faster
than that, because i still forgot to count the big difference between
a windows machine in a tournament hall and a machine on the far
remote running linux.
Definitely wasting all that energy from 8 nodes is kind of stupid
if a dual delivers the same for me! Note this 8 cluster still is
a very cool machine compared to most others.
Now most clusters have huge number of nodes, so the 33% speedup there
will be WAY worse, as they're usually build with 450Mhz or whatever
slow type of processors. This means that on paper a cluster is
not what it looks like in reality.
I wonder which applications in this world can actually take usage
of a cluster (without an smarter rewrite of that algorithm that
runs hell faster on shared memory)!
Best regards,
Vincent
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.