From: "isabel" Newsgroups: comp.parallel.mpi Subject: Re: speedup Date: Fri, 26 Mar 1999 14:06:09 +0100 Organization: Netland Internet Services Message-Id: <922454018.334558@www3.netland.nl> References: <36FA679A.DBCE9106@ecs.soton.ac.uk> Mime-Version: 1.0 Content-Type: multipart/alternative; boundary="----=_NextPart_000_0012_01BE7791.CC778220" Cache-Post-Path: www3.netland.nl!unknown@mx.datasim.nl Xref: ukc comp.parallel.mpi:4849 This is a multi-part message in MIME format. ------=_NextPart_000_0012_01BE7791.CC778220 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Hi, The parallel version of the program doesn't have to be always better = than the sequential version. Sometimes you parallelize the best = sequential algorithm and you don't get the best parallel algorithm. It = might be that your algorithm has a sequential part (part that can not be = parallelizable) very big and a small parallel part.=20 If the amount of computation is not too large, a sequential algorithm = might be faster because the time spent in communication by the parallel = algorithm will be longer than the computation time spent by the = sequential algorithm. Maybe your implementation has a lot of intercomputer communication that = make it very slow. There are a lot of factors that can influence in your = performance. It is not always straight forward to find the right = parallel algorithm. I am sorry I can not give you a better answer. Regards, isabel mohd ezanee rusli wrote in message = <36FA679A.DBCE9106@ecs.soton.ac.uk>... hi everybody.=20 At the moment I'm doing a project to compare the performance of a = parallel algorithm and a sequential algorithm. I'm using MPI for the = parallel version. My question is about speedup. In theory the speedup is = the ratio of the run time of the sequential program to the parallel = program running on P processors. And also the execution time for = parallel version should be less than sequential. However, from my = observation it is not that way. what I see is the execution time for P = processor is more than sequential. Therefore the speedup will decrease = instead of increase.=20 Any idea why?=20 --=20 -Mohd Ezanee Rusli- Computer Science part III,University Of Southampton mer296@ecs.soton.ac.uk. 329 Burgess Road, Swaythling, Southampton UK. SO16 3BA tel: 01703 575399 =20 ------=_NextPart_000_0012_01BE7791.CC778220 Content-Type: text/html; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable

Hi,

The parallel version of the program doesn't have to = be always=20 better than the sequential version. Sometimes you parallelize the best=20 sequential algorithm and you don't get the best parallel algorithm. It = might be=20 that your algorithm has a sequential part (part that can not be = parallelizable)=20 very big and a small parallel part.

If the amount of computation is not too large, a = sequential=20 algorithm might be faster because the time spent in communication by the = parallel algorithm will be longer than the computation time spent by the = sequential algorithm.

Maybe your implementation has a lot of intercomputer = communication that make it very slow. There are a lot of factors that = can=20 influence in your performance. It is not always straight forward to find = the=20 right parallel algorithm.

I am sorry I can not give you a better = answer.

Regards,

isabel

mohd ezanee rusli wrote in message = <36FA679A.DBCE9106@ecs.s= oton.ac.uk>...

hi=20 everybody.
At the moment I'm doing a project to compare the = performance=20 of a parallel algorithm and a sequential algorithm. I'm using MPI = for the=20 parallel version. My question is about speedup. In theory the = speedup is the=20 ratio of the run time of the sequential program to the parallel = program=20 running on P processors. And also the execution time for parallel = version=20 should be less than sequential. However, from my observation it is = not that=20 way. what I see is the execution time for P processor is more than=20 sequential. Therefore the speedup will decrease instead of increase. =
Any=20 idea why?

-- 
-Mohd Ezanee Rusli-
Computer Science part III,University Of Southampton
mer296@ecs.soton.ac.uk.
329 Burgess Road, Swaythling, Southampton UK. SO16 3BA
tel: 01703 575399

------=_NextPart_000_0012_01BE7791.CC778220--