Re: Fastest Matlab server under $4000?
- From: "Colin Reveley" <reveley@xxxxxxxxxxx>
- Date: Tue, 2 Dec 2008 07:04:08 +0000 (UTC)
"Kevin Chen" <hyperq@xxxxxxxxx> wrote in message <gfolru$67r$1@xxxxxxxxxxxxxxxxxx>...
Thanks for the suggestions. I agree that a singler server will be easier to deal with than a cluster.
I checked Dell's server offerings. It looks like I can get a decent machine for a little over $2000. The basic configuration is 2 x Intel Xeon E5420 CPU (Quad core), with 8 (4 x 2GB) ram. The total budget is $4000, so what other hardware should I get? I prefer not to spend another $1000 for CPUs that are just 10% faster.
I am looking for the best performance/cost ratio. Any tips on picking CPUs and RAM to get maximum performance out of Matlab? Should I also look into Nvidia GPUs?
cpp.matlab@xxxxxxxxx wrote in message <a74c2069-2766-4ae5-b98d-f5dbac18cab7@xxxxxxxxxxxxxxxxxxxxxxxxxxx>...
On Nov 13, 9:50=A0pm, "Kevin Chen" <hyp...@xxxxxxxxx> wrote:
I am an research professor building statistical models in Matlab for my r=esearch topics. =A0I recently got a research grant, so I have $4000 (US dol=
lars) to spend on hardware to run Matlab.
a 2 year old, shared server in my school. =A0My question is: What is the f=
My models are very taxing on CPUs. =A0The last one took a month to run on=
astest hardware I can buy with $4000? =A0I simply need to run Matlab and no=
thing else.
cluster of small blade servers? =A0Which one will give me the best perform=
Should I buy one large server with two Intel E5420 CPUs and 8GB ram, or a=
ance/cost ratio?
IF you dont want to change from MATLAB
The best way would be to buy time using that money in Super computing
center at Pittsburgh.
IF you want to change your models
Shift to C++, I had personal experience reducing computational load of
gradient descent method from 4 days to 4 hours
I am not a statistician, so I do not have an intuitive feel for what you might be doing. But, I do have a background in computer science.
It boils down to this: are your problems what are termed 'embarrassingly parallel'?
That is to say, can your problems be broken down into N independent subproblems which can be computed independently? or N problmes that are the same but with different parameters?
a simple example might be a genetic algorithm or other population based search technique. A population of solutions to a problem is generated. Each is assessed at some computational cost.
but as these assessments take place, no communication is necessary between the computational processes. They are completely independent.
If you can frame your problem like that, then the best way to spend 4000 dollars is to buy multiple cheap quadcore systems (e.g. shuttles) to maximise the number of processors you have.
the speed at which your computation progresses will scale nearly linearly with the number of cores.
But what if your problem is not of this nature? what if it is fundametially a serial algorithm?
then you will completely waste your time in buying
"one large server with two Intel E5420 CPUs and 8GB "
because this is a system designed for parallel computing as I have described it. If your algorithms are fundamentally serial, adding CPUs or cores or whatever will not increase speed one iota.
re: matlab v C++, you will find the coding in C or C++ very often produces huge increases per CPU/core. You should look at a beowulf cluster and MPI "'message passing interface". But actually, learning about what matlab does fast (linear operations on matrices) and what it does not (loops and much else) could save you the trouble of redoing it in C.
Say you want lots of computer CPUS. In terms of cost:
Buying 'one large server', even with many cores (especially from sun!!!, or even dell) , is much less economical than buying many small, cheap machines (a small quadcore system can be as little as 400 dollars - so you could get 40 cores for 4000 dollars, plus the cost of networking and visualisation). This is how, e.g. google does business.
If your problem is truly serial, then there is nothing you can do. If it is parallel, you should get large numbers of cheap machines, whether you use matlab or not.
these are trivially the facts.
.
- Follow-Ups:
- Re: Fastest Matlab server under $4000?
- From: Walter Roberson
- Re: Fastest Matlab server under $4000?
- Prev by Date: Re: How to generate whole histogram according to known partial histogram
- Next by Date: Re: fast plotting in matlab
- Previous by thread: Re: How to generate whole histogram according to known partial histogram
- Next by thread: Re: Fastest Matlab server under $4000?
- Index(es):
Relevant Pages
|
