Section: New Results

CAP Bench: A Benchmark Suite for Low-Power Many-Core Processors

Participants : Matheus Souza [PUC Minas] , Pedro Henrique de Mello Morado Penna, Matheus Queiroz [PUC Minas] , Alyson Pereira [UFSC] , Luis Góes [PUC Minas] , Henrique Cota de Freitas [PUC Minas] , Márcio Castro [UFSC] , Philippe Navaux [UFRGS] , Jean-Francois Mehaut.

CAP Bench is an open source benchmark suite that includes parallel applications suitable to evaluate emerging low-power many-core processors such as MPPA-256. The benchmark contains a diverse set of applications that evaluated key aspects of MPPA-256, namely the use of its compute clusters, I/O subsystem, NoC and energy consumption. We expose development difficulties and potential bottlenecks that can stem from the shift in development paradigm when programming for low-power many-core architectures. The results showed us that different applications can have different performance bottlenecks, which is why a solid knowledge about the low-power many-core architecture is necessary for the development of efficient programs.

Our analysis shows that CAP Bench is prepared for the analysis of low-power many-core processors such as the MPPA-256, being scalable and concerned with new trends on this type of architectures. To achieve good performance and scalability, we developed applications considering aspects such as parallel patterns, load balance and architecture limitations. This allowed us to evaluate several aspects of the MPPA-256.

Our benchmark explores the hybrid programming model, which is a trend in low-power many- core processors, following parallel patterns. This enables us to verify that, in the case of MPPA-256, communication time may surpass computation time, which would ideally never occur. This behavior was highlighted by the LU application available in CAP Bench, which may indicate that the NoC should be improved to achieve better performance on NoC-bound applications. In this manner, CAP Bench comes up with the proposal to identify such bottlenecks, revealing potential improvements that might be done in future many-core architectures.

Application development challenges are still out there, and have to be solved to enable the evaluation of next generation many-core processors. As future work, we intend to incorporate other applications to the benchmark, to make it more diverse and allow for a better characterization of the architecture and its aspects. We also intend to extend the benchmark use to other many-core architectures, to achieve a broader understanding of them and the differences between many- core processors

This work was developed in the context of the EnergySFE STIC Amsud project 8.4.2. A description of CAP Bench has been published in the CCPE (Concurrency Computation: Practice and Experience) international journal [9] . CAP Bench will be used and extended during Pedro Henrique Penna's doctoral thesis.