Cpu usage in parrallel running mode

Hi. I have code that running in parrallel mode. I have 4 worker(my cpu has 2 core) but when i run it, it shows that cpu fluctuate from 2 to 24 % but physical memory is in 96%(2×30% , 1×20,2×10%) ,anybody knows why cpu does not show more usage ;like 90%?thank you.

Answer 1

[Walter Roberson]

Have a look at the attached modified version of your code and see if it does what you want. I am not confident that my changes match your intention.

I did some optimizations in the first parfor, but not in the second one (no point doing that until we know that the first part is acceptable.)

... Be careful you do not remove your own good copy of the file before you are sure that my revised version is right!

[ARMIN M]

Ok . Tnx. Do u know how much takes this code to run completely?

[Walter Roberson]

Several hours at least. The time per iteration that I am seeing is about 0.004 seconds but the displays are coming at roughly one second intervals. What that suggests to me is that parfor has a lot of communcations overhead for your purpose and that you would probably be better off not using parfor.

[Walter Roberson]

I found a place that it will fail because of the way that toc() is used. I am running a revised version now and if/when it works I will post it.

[ARMIN M]

Ok thank you a lot. It is 3 days until now that it is running!

[Walter Roberson]

The first phase (the one I made some minor optimizations on) ran for 80 minutes. The second phase (that I did not make the minor optimizations on) ran for 400 minutes. After that the line inv_A=E1*(1/det_AM); failed because E1 is not defined.

[ARMIN M]

I used number to see the answer. But it showes the answer zero at phase1.it is wrong. Whole answer should be 0.2294. det_A22=det_A; E(1,i)=det_A22; S=-1+2*(rem(i+1),2)==0); E1(1,i)=S*det_A22; May need to make preallocation matrix for E1 like E.

[ARMIN M]

I checked the code and found the little mistake in first loop of phase 1. I corrected it and got run. Answer was correct. Tnx. Do you think if i run the code with 4 workers (with 2.4 ghz cpu with 2 core 4 threands and 8gb ram) any problem will occure? Thank you.

[Walter Roberson]

Hyperthreading does not add additional processing threads. Hyperthreading keeps a high speed copy of processor registers for two different processes, and when the first one voluntarily gives up control of the core in order to wait for something to happen, then hyperthreading quickly switches over to the other process and permits it to run until it voluntarily gives up control, and so on.

This configuration does not add any compute capacity. When two processes are both compute-bound, doing all kinds of computations (especially that fit within cache), the effect is to lock out the second process until the first process takes a break. Because hyperthreading adds a little bit to the heat, most CPUs compensate by reducing the clock rate slightly -- and when you do that, the two compute-bound MATLAB routines actually run slower than if hyperthreading were turned off.

[ARMIN M]

I understood that if i use 2 workers it is better. But i used 4 workers before and it was 25 % faster. But technical guide u said illustrate that it may cause slower run because of heat that produced in long time run. that is right?

I have ran code again . U said right . Both of them( 2 and 4 workers) have finished computing in the same time(40minutes). The usage of them were 92 % .but cpu usage of 2 workers decreased to 55% compare to 90% of 4 workers.

[ARMIN M]

I have read your comment. U have edited it. I can not understand finally which configuration is better to use.tnx

[Walter Roberson]

I corrected a minor typing mistake -- I think I had accidentally written "id" in one place where I meant "it".

Using parpool(2) is better since parpool(4) does not improve your performance.

[ARMIN M]

Thank u a lot. So i had understood correct. Sorry two more question 1.i want to simplify the output of this code as much as possible, how can i do this? 2.getting factors of output, takes a lot of time. what can i do to Decrease time of running?

Answer 2

[Nicolas B.]

Hello Armin,

It is already strange that you have 4 workers for a 2 physical cores CPU. Normally, Matlab runs 1 worker per hardware core. So hyperthreading has no impact on the number of workers.

Looking at your use of RAM (96%), I guess that your PC has to use the swap memory which is running on your harddrive and is much more slower. Therefore, if the access to the memory is slow, your CPU will not be used at 100% (bottleneck principle).

If you want to reach the 100% use of your CPU, I may recommand you to add some RAM on your PC. Otherwise, try to reduce the number of workers.

[ARMIN M]

2 core .4 threads .In parpool , matlab accepted 4 workers(parpools(4)).at first my laptop,s ram was 4gb,i upgraded to 8 gb and my laptop(motherboard) does not support more. Reduce workers to 3 or 2 make it better?thank u.

[Nicolas B.]

On my PC, I consider that, to run MATLAB, I need:

• 2 GB for the main matlab (the process with the window)
• 600 MB per instance on parallel toolbox

So if I do the maths, for 4 workers, I need 4.5 GB of RAM. If you have 8 GB of total RAM, it's a lot.

Maybe, you can provide an example of your code and I can try to see if the problem come from your PC or your code.

[ARMIN M]

This is my code. 1 problem occurred during get running from code. Parpool had shuteddown automatically. Do you think it is possible? Or is it possible to shutdown manually from down corner of matlab window during code runes? Tnx.

[Walter Roberson]

Very little of your code is being done in parallel. Only lines 234-248 and 357-372

[ARMIN M]

Yes . I did not know about parallel running and i could just learn this in 2 ,3 days. I should compute this code soon and i don,t have time to learn how to make it better.

[Walter Roberson]

Near line 256 you calculate

det_A=sum(H,2)

This is the last place you use H, so if the value of this det_A is not used then you have no reason to calculate H.

Then you do

                F(1,iiii)=a4*det(A_4);

which does not use det_A . You do not use det_A anywhere inside that loop.

You

After the loop, you do some more calculations to build the rest of F, and assign into det_A based on F -- the previous det_A going unused.

[ARMIN M]

Hello walter .thank u . Yes .you say right. It should be F(1,iiii)=a4*det_A, actually this code calculated det(A) twice. Which takes more time to complete. thank you.