1

My current computer has 16 cores and 32 Logical processors. I am using Matlab 2014a. The upper limit for number of workers to use is 512. In my case, what is the optimal number to use?

If it takes, say 20 minutes, to finish a job under 

    matlabpool open 16

for the same job how long would it under 32 workers? 10 minutes or more?

Jie Wei
  • 191
  • 9
  • Also read [this answer](http://stackoverflow.com/questions/32146555/saving-time-and-memory-using-parfor-in-matlab/32146700#32146700) for a more general overview for the mechanics of `parfor` and what the overhead does. – Adriaan Jan 03 '16 at 17:43
  • In general: Daniel's answer is a good start, but without seeing your code we can't say anything. Usual strategy is optimise as far as possible in your serial code, then go parallel and see whether there's any improvement, taking Daniel's points into account. – Adriaan Jan 03 '16 at 18:06

2 Answers2

2

The number of workers depends on multiple factors.

First of all, how many cores can a worker use. Let's assume a typical computation intense task where your workers get large chunks of work. In such a scenario you can assume each worker to occupy at least one physical core. Don't use more workers than physical cores available.

Further, when your code is heavily based on multithreading enabled functions, each worker can use more than one core. Especially when using one of the mentioned functions with large images or matrices, you can expect the best performance with much fewer workers. With an octacore system, I ended up not using the parallel computing toolbox at all because a single matlab thread already used the full CPU capacity, only adding unnecessary communication overhead for the parallel computing toolbox. I assume such scenarios have lead to the recommendation of at most 1 worker per CPU in clusters.

Besides your CPU capacity, take a close look on the memory usage. Regardless of the CPU capacity, as soon as your system starts swapping, the performance dies. I recommend to reduce the number of workers as soon as you observe peak memory usage with leaves less than 1GB free memory.

When your workers load or write data to the HDD, also take a look at it. The best metric for HDD usage is the queue length, as soon as it never reaches zero, you can assume your HDD to be in full usage. More workers won't speed up anything.

If I would have to make a blind guess, I would try it without using the parallel computing toolbox and with 8 workers.

Daniel
  • 36,610
  • 3
  • 36
  • 69
  • Hi Daniel, thanks for your kind reply. I just did a simple comparison by running the same task by using 12 and 16 workers (in turn). It seems 16 workers use about 90% of time that 12 workers do. And when 16 workers are running, I see my "CPU usage" is 53% and "memory usage" is 16G(out of 64G). So just wondering if my computer is capable doing more workers. Thank you. – Jie Wei Jan 04 '16 at 07:52
  • With 53 percentage CPU usage, it seems something is permanently blocking your computation. This might be because your workers spending to mouch time communicating instead of doing computations. In this case it might be faster with more workers, but optimizing the code to avoid the blocking would probably increase the performance. – Daniel Jan 04 '16 at 12:41
0

the short answer is that it is highly variable, dependent on your hardware, your code implementation and the nature of the problem itself.

in general however, a linear speed (ie, ten minutes reduces to five by doubling number of processors) up is the very best case scenario. in reality there is some of amount of overhead associated with parallelism.

your question is open ended. I suggest reading matlab documentation on efficient parallelism strategies

crowdedComputeeer
  • 469
  • 2
  • 6