ИСТИНА

While opening a lot of exciting possibilities, the HPC systems with heterogeneous architectures (CPU+GPU) have posed two new challenges in software development, namely (i) how to port the existing software to heterogeneous world or to develop it from scratch, and (ii) how to efficiently optimize the resulting application to get a pronounced performance gain for a given HPC platform and particular data. Today, there are several solutions on the market that facilitate the resolution of the first issue. However, the second problem so far remains in effect unanswered. In this address, we will discuss the major problems of software optimization on heterogeneous platforms and present our practical experience on getting them around.