A flexible shared library profiler for early estimation of performance gains in heterogeneous systems

The effective acceleration of computationally demanding applications in heterogeneous systems often requires significant optimization efforts. Although such task typically starts with a thorough profiling stage, a special attention must be given to the migration procedure of each application kernel: apart from the actual computation time, the cost of the data transfers between the main processor memory and the accelerator plays a significant role, which often limits the actual resulting speedup. In some cases, no performance gain is actually achieved, given the excessively high communication to computation ratio. To ease the system designer effort, this paper proposes a framework that transparently collects extensive profile information, including, but not limited to, the values of the processor performance counters, as well as an estimation of the amounts of data to be transferred to and from the accelerator. The framework focuses on transparent acceleration of kernels implemented as library functions and is based on the shared library interposing technique. By further processing of the obtained execution profiles, together with the proper communication and computation models, the attainable global speedup of the accelerated application is predicted. The presented methods were validated experimentally for a set of existing applications. The measured global speedup estimation error typically ranged between 1 and 4%.