Dear Editor Tim Curns,
I have come across a copy (http://www.tgc.com/hpcwire/hpcwireWWW/05/0204/109182.html) of your interview with Mr. Kevin Howard, the President and Co-Founder of Massively Parallel Technologies, Inc. I am writing concerning a statement made by Mr. Howard:
“… Companies invest millions of dollars in these systems yet only experience a maximum of 20% efficiency over a single processor. Furthermore, as the number of nodes and file sizes increase, efficiencies drop to well under 10%.”
Although I am not defending companies that invest millions of dollars in supercomputing systems, I am uncomfortable with Mr. Howard's statement of “.. as the number of nodes and file sizes increase, efficiencies drop to well under 10%.” I have developed many algorithms for several years on both shared and distributed architecture computers. Also, I have been developing cluster computing systems with Windows XP operating system and with wireless network connections. I believe my experiences show otherwise. As a result, I strongly disagree with Mr. Howard. His statement is especially misleading if the file size increases proportionally to the increase of processors.
I believe Mr. Howard might have meant to state : as the number of nodes increases and the file sizes (amount of work per node as the number of nodes increase) decreases, efficiencies drop to well under 10%. As is known, parallel efficiencies of problems are compared with those of serial counterparts. I believe if Mr. Howard meant to say what was published, then the problems being solved on one-node and multi-node are not comparable. Therefore, there should not be any comparison on the performance (or parallel efficiency) of such problems.
I would like to know Mr. Howard's further clarification on his statement.
Sincerely yours,
Musa Yavuz
President and Founder,
United Sciences Research Group, Inc.