Re: Help! Mathematica on my 500MHz outperforms my GHz machine!
- To: mathgroup at smc.vnet.net
- Subject: [mg23049] Re: Help! Mathematica on my 500MHz outperforms my GHz machine!
- From: terryis at my-deja.com
- Date: Thu, 13 Apr 2000 02:43:32 -0400 (EDT)
- Sender: owner-wri-mathgroup at wolfram.com
Fellow Mathematica Users, After spending some time investigating the lackluster performance of my gigahertz Athlon machine I have finally discovered why it only performs at about the same level as my 500MHz Pentium III. In short my Athlon is bottle- necking on the relatively slow 100MHz bus between cache and main memory (as some of you had suspected). To start with I ran the various benchmarks suggested to me. From http://fampm201.tu-graz.ac.at/karl/mma.html I ran the MMA-Test.nb (for Mathematica 3.0) and got the following results: Version = Microsoft Windows 3.0 (April 25, 1997) Times = {0.531,0.591,0.18,0.371,1.91,0.661,0.181,0.28,0.551,0.641,0.47,0.862,0.5 81,0.49,0.461} Total Kernel Time = 8.763, Total Time = 9. Benchmark = 11.2 The benchmarks at www.scientificweb.com/sciencee/mathstef2.html could not be found and the ones at www.geocities.com/SiliconValley/Lab/2255/bytemark.html worked on one of my computers but gave a 'numeric sort' error condition on the gigahertz machine. The link 205.181.113.18/zdbop/zdbop2.html could not be contacted despite multiple attempts on different occasions. Next I split up my simulation and timed each piece individually on each machine. For each piece my Athlon kicked ass even though it performed poorly on the whole simulation. Finally I was able to localize the difference in performance to a single piece of code that saved intermediate results to memory. With the code to save intermediate results in, my Athlon took on average 27 minutes to run 25000 iterations of my simulation compared to 24 minutes on the Pentium. Without the code to save intermediate results the Athlon was averaging 1 minute compared to 3 minutes on my Pentium! My conclusion was that without saving intermediate results, the sum total of all the parts of my simulation was able to fit completely in cache and so could run at the full 200MHz bus speed between cache and CPU. Interestingly enough the performance of the Pentium also improved dramatically (from 24 minutes down to an average of 3 minutes!) which to me indicated that the cache hit ratio on the Pentium was not that high either. As it turns out, my simulations can be re-written to take full advantage of this machine. Even if I have to go back and re-run segments of my simulation to re-capture those intermediate results that are worth capturing, it's still worth it for me. Bottom line: if you are certain that what you are doing can fit into cache then this could very well be one of the faster machines for you to run Mathematica on. Otherwise I'd stick with a Pentium based machine. Finally check out www.zdnet.com/pcmag/stories/firstlooks/0,6763,2426421,00.html for a good article that sums up this machines deficiencies. Sent via Deja.com http://www.deja.com/ Before you buy.