HPC as a Service in Education

Abstract

Advances in Cloud Computing have opened many chapters in Information Technology. Numerous service platforms offer clients of the cloud ease of use and flexibility of using the provided services. Education with billions of potential users worldwide is a major target. An emerging service called HPC-as-a-Service (HPCaaS) targets Science, Technology, Engineering, and Math (STEM) users. In this paper we discuss the use of HPCaaS platform in STEM education. We argue that such a service can significantly alleviate a major obstacle in teaching parallel programming for the STEM students. Cloud computing provides unique benefits such as resource pooling, cost efficiency, availability, and large computational power. These features have attracted scientists, engineers, scholars, and the High Performance Computing (HPC) users like a magnet towards the Cloud. However, HPC programs often consume large number of collaborating processors to reduce the execution time, where synchronization between these processors and the communication overhead among them can become real challenge even on dedicated and special hardware, but worse on shared and virtualized platform like cloud. As a result, moving HPC applications to the cloud can adversely impact the abovenamed difficulties with additional issues, primarily virtualization, multitenancy, and network latency. One solution is a new cloud service known as HPC-as-a-Service. In this paper, we present an HPCaaS platform called ASETS which uses Software Defined Networking (SDN) technologies to smooth the execution of parallel tasks in the cloud. Further, we provide application examples that could be used in a typical introductory parallel programing course. We argue that HPCaaS platform like ASETS can significantly benefit the users of HPC in the cloud as if their program is running on a dedicated hardware in their own laboratory. This is especially advantageous for students and educators who need not to deal with the underlying complexities of the cloud.

Improved Matrix Multiplication by Changing Loop Order

“…It is worth noting that not every program can be converted in parallel, and there are cases where the parallel version exhibits poor performance compared to the serial version. With the latest computing architecture and parallel languages, tremendous improvements have been reported for machine learning, AI, a lightweight matrix-based key management protocol for IoT networks, graphics, computational photography, and computer vision by exploiting parallelization [3][4][5][6][7].…”

“…In this paper, we limit our analysis to a multicore system by observing the program behavior and changing loop orderings [4][5][6][7]. We apply a parallel construct for matrix multiplication to gain better performance.…”