Optimized Data Transfers Based on the OpenCL Event Management Mechanism

Figure 2

Overlapping communications and computations. (a) The communication time is overlapped with the computation time. (b) The computation time is too short to hide the communication time. Since joint programming of OpenCL and MPI cannot express the dependency between the first communication and the second computation, the host thread is blocked to execute them in a correct order. (c) The second communication can potentially start earlier because the host thread is not blocked.

(a)

(b)

(c)