Mathematics performance and hardware

[Carlis]

Mostly out of curiosity;
What are the critical hardware resources for good performance under math programs like mathematica and matlab... Is it ram or processor speed? Does such processes make use of many cores?

 

[trOver]

sorry, but those programs arnt the most popular for benchmarking, so theres not much on them. I would say that most mathematics is going to be cpu work, because ram is just used for accessing the hd and a temporary storage space for the cache of the cpu. Because there is not much graphic work involved, i would say that cpu speed would be the best things in making your decision for running those programs

 

[MagnusTheBrewer]

Depending on the level of accuracy required, I could see a case being made for registered ram but, unless you moonlight for NASA, any system capable of running the software will give you simular performance. So, break out that 286 processor and have some fun! 🙂

 

[Matthias99]

Originally posted by: trOver
sorry, but those programs arnt the most popular for benchmarking, so theres not much on them. I would say that most mathematics is going to be cpu work, because ram is just used for accessing the hd and a temporary storage space for the cache of the cpu. Because there is not much graphic work involved, i would say that cpu speed would be the best things in making your decision for running those programs

"ram is just used for accessing the hd"? I take it you don't do a lot of programming.

Number-crunching performance depends wildly on the sort of computations you are doing. Some algorithms require massive amounts of (fast) integer math -- frequently the limitation here will be the RAM, because the datasets are generally far too big to fit in the CPU's L1/L2 cache and so it spends a lot of cycles waiting to fetch data or write results. Others require a lot of floating-point operations... depending on the exact CPU, and what operations you are doing, this could end up being limited by the speed of the FPU more than anything else. Some operations (usually ones that do relatively simple things to large amounts of data) can benefit from deep CPU pipelining -- but algorithms that constantly have to branch and make decisions based on their results tend to get hurt by overly deep pipelines, since they will frequently encounter pipeline stalls.

Modern x86 CPUs also offer specialized instruction sets (MMX, SSE/SSE2/SSE3) for doing SIMD (Single Instruction Multiple Data) or 'vector' operations on large amounts of data in parallel. If you have to do the same exact integer math to a vast quantity of data, using these instructions can increase performance significantly. Programs like Matlab would usually offer some method of accessing this functionality -- consult your documentation or an expert user.

I'm pretty sure that neither Mathematica nor Matlab is by default multithreaded (unless this is a recent development -- Matlab definitely wasn't a few years ago). There may be packages you can access that would allow it to use multiple cores for some operations -- or you could run multiple instances of the program.

 

[Aluvus]

Originally posted by: MagnusTheBrewer
unless you moonlight for NASA, any system capable of running the software will give you simular performance. So, break out that 286 processor and have some fun! 🙂

Erm, uh, no. I don't have a great deal of experience with Mathematica, but I know Matlab likes to have a pretty powerful system even for fairly simple stuff.

Matthias99 is right though, that the answer to the OP's question will depend largely on the nature of the calculations being done. My general experience with Matlab has been that more RAM is valuable (particularly when dealing with large data sets, as one would expect), but once you have a reasonable amount of RAM processor speed becomes the major concern. Again, this is just my experience, based on the sort of (mostly not all that intensive) use I'm familiar with.

Matlab is single-threaded, though at various points it has been able to connect to libraries that make use of multithreading. Mathematica can apparently multithread some operations as of version 5.

 

[MattM76]

A little different- but I know when I doing phylogenetic calculations using PAUP/Arlequin my CPU pings up to 50% on my e6600. Seems to indicate processor importance.

Matt
STD Blog

 

[StormRider]

I think it depends on the type of mathematical processing you are doing. Numerical computation (like in MatLab) probably depends more on the FPU while symbolic computation (Mathematica) probably depends on integer performance. Mathematica can do both symbolic and numeric processing....

 

[Carlis]

Yes matlab is painfully slow sometimes, I have been told its written in java. It is the default algebra program on my university. Many rather simple calculations can take 15-30 min(sparc´s & 1.6ghz amd) and soon I expect dealing with differential equations that will be far more messy.
From what I have heard mathematica is significantly faster and contain better graphic abilities. Maybe thats the way to go.

I just got my self a C2D 2.16 ghz 1 gig RAM laptop. I havent run much math on it yet but hopefylly that will fix it.