[Date Index] [Thread Index] [Author Index]
Re: Mathematica and 64-bit computing (... and Unix/Linux)
We have a Dec AlphaServer 2100 (275MHz, I think, with tons of RAM) with Digital Unix 4 and Mathematica 3.02. I have always wondered why Mathematica runs so slowly on it. It is slower than the same notebooks on an Apple Powerbook 3400c at 200MHz with 80MB RAM. Doesn't seem to matter much if it is basically floating point computations or symbolic algebra manipulations. I have some hopes that Mathematica 4 will take more advantage of the 64 bit architecture. You may have noticed that Mathematica 3 came out for the DecAlpha a year or so after it came out for other machines. However, this time a Wolfram rep has told me that 4.0 is coming out now with the others. The reason he gave is that they had a lot of issues initially with the 64 bit architecture which were solved during the version 3 port. He also says that the demand for Mathematica on Mac's and PC's dwarfs that for Unix versions. Also, I have heard that Mathematica runs similarly slowly on SGI's and Sparc's (compared to what you would expect from their floating point and integer benchmarks). The truth may be that Wolfram just spends more effort on optimizing for Mac's and PC's relative to Unix. Further evidence of this may be the exception to the rule -- Mathematica runs fine on Linux, even though it is just another type of Unix. A possible reason ? -- Wolfram does most of their basic development on Linux (according to the sales rep)! This suggests yet another theory. I was reading some Linux news somewhere that Linux on a Sparc beats Solaris on the same Sparc machine in some applications where only a single processor is used. Apparently, there may be some sacrifice in single processor performance built into the commercial Unix's operating system for the benefit of their big multiprocessor applications. Could this be another possible explanation why Mathematica on Linux is fine, but not so great on the DecAlpha? It should be obvious the above is speculation. Anyone have better info? Guilherme Roschke wrote: > > Well were currently running version 3 on a RISC/6000. What we are > interested in is improving our floating point performance, and have > already noted the benefits of Version 4 on that field. > > I have also noted the great SPEC_fp benchmark numbers that compaq's > (formerly digital) 64-bit alpha processors have. I'm guessing this is > due to their 64 bit architecture, and that these benchmarks were > compiled native to that. > Mathematica (versions 3 and 4 ) is available for Tru64 (formerly > Digital Unix). What i'm wondering is exactly how much of this > architecture does mathematica take advantage of? > > Does it improve the $MachinePrecision? currently i'm at 16. > > Does it do its floating point taking advantage of the local > architecture? i.e. (nearly) double the speed? > > I noticed the packed array stuff only works with machine precision (not > arbitrary) so an improvement in machine precision would still be nice.