New discoveries in science and engineering are driven by simulations on high performance computers - especially when physical experiments would be unfeasible or impossible. Milinda Shayamal Fernando's research is focused on developing algorithms and computational codes that enable the effective use of modern supercomputers by scientists working in many disciplines.
His key objectives include making computer simulations on high performance computers: easy to use - by using symbolic interfaces and autonomous code generation; portable - so they can be run across different computer architectures; high-performing - so that they make efficient use of computing resources; and scalable - so that they can solve larger problems on next-generation machines.
Milinda Shayamal Fernando's work has enabled improved applications in areas of computational relativity and gravitational wave (GW) astronomy. In the universe, when two supermassive black holes merge, they bring along corresponding clouds of stars, gas and dark matter. Modelling these events requires powerful computational tools that consider all the physical effects of such a merger. While recent algorithms and codes to develop simulations of black hole mergers have been developed, they were limited because they could only handle simulations when the masses of the two black holes were comparable. Milinda Shayamal Fernando developed algorithms and code for mergers of black holes, or neutron stars, of vastly different mass ratios. These computational simulations help scientists understand the early universe as well as what is going on at the heart of galaxies.
A general problem in high performance computing occurs when multiple distinct jobs running on supercomputers send messages at the same time, and these messages interfere with each other. This inter-job interference can significantly degrade performance.
Staci Smith's first research paper in this area, "Mitigating Inter-Job Interference Using Adaptive Flow-Aware Routing", received a Best Student Paper nomination at SC18, the premiere supercomputing conference. Her paper had two goals: to explore the causes of network interference between jobs - in order to model that interference - and to develop a mitigation strategy to alleviate the interference.
As a result of this work, Staci Smith recently developed a new routing algorithm for fat-tree interconnects called Adaptive Flow-Aware Routing (AFAR), which improves execution time up to 46% when compared to other default routing algorithms. As part of her ongoing PhD research, she continues to develop algorithms to improve the performance and efficiency of HPC workloads.