Messed up timing information

Accelerated Computing CUDA CUDA Programming and Performance
1

I have been trying to get timing information using the cutcreatetimer(). I have been running a loop multiple times in order to get accurate timing results. Now for just 1 iteration I get a timing info as under:

bash-3.2$ …/…/bin/linux/release/./Izhikevich 24 1
Processing time: 0.305000 (ms)
-bash-3.2$ …/…/bin/linux/release/./Izhikevich 24 2
Processing time: 0.458000 (ms)
-bash-3.2$ …/…/bin/linux/release/./Izhikevich 24 10
Processing time: 1.705000 (ms)
-bash-3.2$ …/…/bin/linux/release/./Izhikevich 24 100
Processing time: 271.239990 (ms)

first argument is the image size: 24 in this case and the next one is the number of iterations. For instance in the last run, it takes 271ms for 100 iterations and hence 2.71ms for one iteration. It is different from other runs. Can any one suggest why the discrepancy is there?

The loop involves device host transfers and kernel call.

Regards

Vivek

2

first do you remove warmup time?

second you cannot trust sub-mini second in one run (it has large variation),

you should do average, so the better way is to compare

-bash-3.2$ …/…/bin/linux/release/./Izhikevich 24 100

and

-bash-3.2$ …/…/bin/linux/release/./Izhikevich 24 1000

or try large image size