Calculation Acceleration

Question

I have a modeling application that I wrote in both C and CUDA. To measure acceleration, I recorded time in both cases. In CUDA, I used CUDA events to measure the time and then divide the GPU time by the CPU (as is usually done). The acceleration image is shown below.

The strange thing about the acceleration graph is that the acceleration first increases to 55X, then decreases to 35X, and then increases again as the total number of threads increases. I am not sure why this is happening and how I can find out the reason for this conclusion. I am using a GTX 560ti GPU card with 448 cores. The number of threads for each block is 1024 (maximum number) and therefore one block for each SM. This is due to employment problems and how can I definitely find out the reason for such an acceleration schedule?

Answer 1

Acceleration peaks seem to be related to CPU runtime. By analyzing GPU time, it seems to increase linearity with the number of agents. However, CPU time, which also increases linearity in general terms, has a fall time in the range [0.6,1.6] aprox, and some peaks in the range [2.6,3.1] aprox.

Considering the above, your maximum acceleration is 55 times reduced in the range [0,6,1,1] aprox. because your cpu time is also decreasing. Therefore, to calculate acceleration as CPU time / GPU time is normal that the result is smaller. The same applies to the second, in the range [2.6,3.1] .

How can I find out the reason for this acceleration graph? I assume that the processor was interrupted by some external event (I / O, another program running on the CPU, OS ...).

To more accurately calculate accelerations, repeat the experiment 10 times as individual executions , i.e. don't create a loop inside your main function to execute it 10 times. Using 10, 20, 30 or even more individual designs, you can calculate the average time as well as the variance . Then, the time taken to complete the study: one or two peaks can be considered as special cases (ignore them). If you see a trend, then a deeper study is needed.