what does Visual Profiler mean regarding this analysis result? Kernel time+Memory copy time does not

Accelerated Computing CUDA CUDA Programming and Performance
1

platform: vs2010+tesla c2050

i run a test program on tesla c2050, and use visual profiler to analyze the program. the program have 3 memory copies between cpu and gpu and 1 kernel function. Visual Profiler analysis result are as follows:

Kernel time = 0.51 % of total GPU time
Memory copy time = 36.2 % of total GPU time
Kernel taking maximum time = VecAdd (0.5% of total GPU time)
Memory copy taking maximum time = memcpyHtoD (25.0% of total GPU time)
There is no time overlap between memory copies and kernels on GPU

my question is, why the Kernel time+Memory copy time does not equal 100% of GPU time?

2

The total GPU time is computed as the difference between the end of the last kernel or memory copy and the start of the first kernel or memory copy.

The kernel + memory copy time may not add up to 100% of the total gpu time - due to gaps on the timeline when no kernel or memory copy is executing. Such gaps can be due to some cpu side processing or other overheads such as API or synchronization overheads.