problem with the executed time

[Anthony]

[sebraa]

You do not measure the time to run your program on the Epiphany.
Instead, you measure the time it takes to copy your program into the Epiphany. Of course, copying the code into 16 cores takes about 16x as long as copying it to one core only, since there is 16x as much data to copy.

You cannot measure the Epiphany execution time on the host directly; usually, you keep a flag in shared memory which you poll in a loop. Then you measure the time from after the e_load_group() until this flag is set by your core.

[Anthony]

[sebraa]

Accessing the Epiphany memory from the host or shared memory from the eCores is an access through the eLink. If you poll by reading Epiphany memory, you waste a lot of bandwidth on the eLink interface, which might or might not affect your timing.

[Anthony]

Hi,sebraa!
i agree with your opinion,but i have to get the time to caculate the speed of my algorithm ,do you have any idea about this,so that i can get the executed time correctly as far as possible

Best Regards

Anthony

[sebraa]

Well, first:
- don't spam the forum with similar questions.

Then:
- read the documentation.

You will notice that the eCores contain two timer units each, which you can use to measure clock cycles on the Epiphany. This will provide you with "Epiphany algorithm execution time", instead of "sending program and waiting for result wall clock time", which you get with gettimeofday() on the host.

It depends on what you actually want to do. Judging from your current posts here, you are not sure about this yourself. So, take a step back, and think carefully about what you actually need.

Also, measuring something always comes with overhead, which will affect your measurements. Even if you don't care, you should at least know about it (and maybe give a reason why this doesn't invalidate your measurements).