[Bioclusters] Related question (was: a dedicated cluster to mpiblast the nr database)

[Richard Walsh]

Joseph Landman wrote:

>This raises an interesting question ...
>
>Tim Cutts wrote:
>
>[...]
>
>>We've done this recently, comparing 32-bit BLAST on Linux/x86, with 4GB
>>RAM, against various vendors' quad-CPU-64-bit-32GB-RAM performance
>>monsters.  For BLAST, in particular, I don't think big machines are
>>worth it.  The 20-40% speed advantage is much smaller than the price
>>penalty.  Buy lots of little ones instead.  If you need a more generic
>>  
>>
>
>I am curious about this.  I agree with Tim on this, but I have not been 
>able to quantify this for myself, even from an estimation point of view.

 Paying the design price to consolidate hardware (SMP, memory subsystem,
 etc.) is primarily beneficial when there is a many-to-many interaction 
 between program instructions and program data (for the bulk of the apps
 instructions, data may be required from any of a large number of memory
 /remote locations).  Problems which do not have this feature get by in 
 a less integrated "share-none" environment with the added and complicating
 benfits of:

           * super linear speed ups from global increases 
             in cache, memory, and bandwidth.

           * drops in the price per register-resident operation 
             as lower cost processors are selected.

>What speed advantage would be worth such a price penalty?  Or better 
>put,  for every $1000USD over in price per unit from the base, how much 
>additional performance would be needed to justify the cost?   Is the 
>idea to keep the price performance the same (e.g. for 10x the price, you 
>get 10x the performance), or should it be a different function.

 It would seem that at a fix point in a Moore's-law-neutral amount of
 time, the buyer would >>like<< 10x the price to deliver 10x the performance;
 but from an engineer point of view this clearly cannot be delivered
 in all performance regimes.  As the seller travels up the performance
 curve, they hope to find buyers with absolute capability requirements and
 the cash to match.  The market becomes discontinuous at these places
 (like the market for expensive homes) and the combination of market
 and technological risk limit the size of the performance improvement
 bite any seller willl take in designing the next level of performance 
 into their product. 

>Another similar question is, if these larger boxen were 1/10 their 
>current price, would people buy 10x more?  Less than that?  More than 
>that?  That is, is this the analysis bottleneck?

 It would seem that they would by 10x only if they actually could 
 consume 10x the capacity. If apples drop in price by 10 times, we 
 would only buy 10 times as many if we could/wanted to eat ten times as
 many apple pies ... ;-). When the result is delivered to another consumer,
 the rate of use/consumption at their end of the 10x capacity product 
 is the key issue. This suggests the relevance of rate of use issues and 
 bootleneck shifting typical in any production line.  I will let you 
 continue with the apple pie metaphor ... ;-), but the demand for life
 saving drugs is certainly as that for a good apple pie.

 ;-)

 rbw
#---------------------------------------------------
# Richard Walsh
# Project Manager, Cluster Computing, Computational
#                  Chemistry and Finance
# netASPx, Inc.
# 1200 Washington Ave. So.
# Minneapolis, MN 55415
# VOX:    612-337-3467
# FAX:    612-337-3400
# EMAIL:  rbw@networkcs.com, richard.walsh@netaspx.com
#         rbw@ahpcrc.org
#
#---------------------------------------------------
# "What you can do, or dream you can, begin it;
#  Boldness has genius, power, and magic in it."
#                                  -Goethe
#---------------------------------------------------
# "Without mystery, there can be no authority."
#                                  -Charles DeGaulle
#---------------------------------------------------
# "Why waste time learning when ignornace is
#  instantaneous?"                 -Thomas Hobbes
#---------------------------------------------------