Stream-based Parallel Computing Methodology and Development Environment for High Performance Manycore Accelerators

The latest supercomputers incorporate a high number of compute units under the form of manycore accelerators. Such accelerators, like GPUs, have integrated processors where a massively high number of threads, in the order of thousands, execute concurrently. Compared to single-CPU throughput performance, they offer higher levels of parallelism. Therefore, they represent an indispensable technology in the new era ∗E-mail address: [email protected] †E-mail address: [email protected] ‡E-mail address: [email protected] §E-mail address: [email protected] 2 S. Yamagiwa, G. Falcao, K. Wada and L. Sousa of high performance supercomputing. However, the accelerator is equipped via the peripheral bus of the host CPU, which inevitably creates communication overheads when exchanging programs and data between the CPU and the accelerator. Also, we need to develop both programs to run on these processors that have distinct architectures. To make it simpler for the programmer to use the accelerator and exploit its potential throughput performance, this chapter describes the Caravela platform. Caravela provides a simple programming interface that overcomes the difficulty of developing and running parallel kernels not only on single but also on multiple manycore accelerators. This chapter describes parallel programming techniques and methods applied to a variety of research test case scenarios.