Recommend the best possible machine for the price.

I've tried searching through other Q&A's, but haven't found a definite answer since everyone's use of Matlab is different.

Can anyone recommend an HPC that will give me the best bang for the buck around a $14K budget? Looking for a Windows machine in desktop or rackmount (5U max) format. We have access to all Matlab toolboxes.

A little background on the types of matlab analysis our research group does...

We have millions of JPEG images of open ocean that are sized around 3-5Mb each and we do thresholding to compute white cap size coverage. The guy that does this analysis is currently using the following... Computer Model: XPS 8300 Intel Core i7-2600 processor(8MB Cache, 3.4GHz) RAM: 8GB DDR3 SDRAM at 1333MHz - 4x2GB OS: Windows 7 (64Bit) Video Card: 1GB ATI Radeon HD 5670 I believe he has said that at the current rate of analysis, it'll take a couple of years of processing to get through all the images. If we can get that down by a factor of at least 10, that would be great.

Other things we do are fft's on data arrays and other run of the mill processing on large ASCII files. I should mention that we worked with NVIDIA's cluster to test some of our fft code on one of their top of the line GPU enabled machines running the latest and greatest Tesla GPUs and found that the GPUs didn't help very much. (We weren't able to test our image processing code on the NVIDIA machines so the GPUs may or may not have helped in that regard. My guess is that the GPUs can probably help tremendously processing images in parallel, but I'm not 100% sure.)

I believe it is the way our data is structured that it can't take advantage of the way GPUs work. (Our data arrays may not have been large enough.) In any case, we came to the conclusion that the best machine for our needs will have as many fast processing cores and a large amount of RAM that we can afford. We will probably get a GPU in there as well just to have one for future needs.

As for the $14K budget, we are not opposed to paying an extra few thousand if it means we are going to get much more processing power or memory.

Does Mathworks do in-house benchmark testing on specific brands of machines like Dell or other name brand HPCs? If so, what are the results? If anyone out there can recommend a system or point me in the right direction, that would be great. Thanks.

[Daniel Shub]

Can you provide some details about what the bottle neck is. You mention in the question and comments that you have millions of images. Is each image (or maybe groups of images) processed independently from the others?

Answer 1

[Daniel Shub]

It sounds like your problem, or at least a large chunk of it, is embarrassingly parallel. If you really can do the analysis with the current computer, but just need a long time, I have two suggestions.

The first is to get some time on a compute cluster (either pay, trade results, collaborate, etc). With the current power of compute clusters, you could reasonably reduce your compute time by a factor of 1000.

The second is to buy $14,000 worth of "cheap computers". You can probably get a computer close to your current specs for $1400, so that will be your factor of 10 reduction. It won't fit in your 5U requirement and will have greater power consumption.

[James]

We did try paralleling the code on NVIDIAs machine, but it didn't help. So I don't think it is embarrassingly parallel. If you are referring to the image processing part, then yes, we may have to do more testing on that end.

"It won't fit in your 5U requirement and will have greater power consumption." --- Exactly why we can't do what you are suggesting.

[Walter Roberson]

It would also need a lot of licenses.

If you are going to get more than 12 cores, the software costs are going to use up your budget pretty quickly.

[James]

Can you point me to a vendor from which you are basing your comment? We are looking for a fast machine, but reliability is also a major selling point. I've looked around and there are a bunch of HPC sites out there and would like to get an idea as to what is dependable and what isn't.

[Daniel Shub]

@Walter, for a lot of things you can use compiled code, meaning you only need one license of MATLAB and the compiler.

[Walter Roberson]

@James:

SGI (sgi.com) does a lot of blade and cluster work. I do not, however, have any idea what their price-points are these days. (I worked with SGI equipment in their IRIX MIPS days, but not since then.)

@Daniel:

http://www.mathworks.com/products/compiler/compiler_support.html

"Executables and shared libraries built from parallel MATLAB applications that use Parallel Computing Toolbox functions require MATLAB Distributed Computing Server to perform computations on a cluster. These executables and libraries run serially in standalone mode in absence of the server product."

But Yes, you could put together your own scheduler of independent (serial) processes, each set to use only one thread (e.g., to avoid clashing with each other as each asks for multiple cores do do LAPACK or the like.)

The Compiler itself is about $5K, which is a third of the budget already.

Answer 2

[Jan]

Buying fast hardware cannot have the same effect as paying a high-skilled programmer. There are so many pitfalls in Matlab and even more, when parallel programming is required.

10'000$ can yield a speedup factor of 2 or 4, when invested in hardware. If you want a factor of 10 or 1000, buy a programmer.

[James]

Easier (and cheaper) said than done. $14K doesn't buy a programmer for very long.

[James]

Can anyone recommend trustworthy vendors or comment either way if using a GPU can significantly speed up image processing of millions of images?

Answer 3

[James]

Does Mathworks do in-house benchmark testing on specific brands of machines like Dell or other name brand HPCs? If so, what are the results? Are there specific vendors that Mathworks likes to use or can recommend?

[Walter Roberson]

MathWorks only has that in one regard (that I know of): it offers turnkey solutions for prototyping and HIL. http://www.mathworks.com/company/pressroom/MathWorks-Launches-Turnkey-Solution-for-Rapid-Control-Prototyping-and-Hardware-in-the-Loop-HIL-Simulation.html

liquidnitrogenoverclocking.com routinely does MATLAB benchmarking on its systems. Like their 5 GHz i7 systems. Which, unfortunately, are not rack-mounted.

accelereyes.com has GPU benchmarks, I believe.