The upcoming Intel Knights Landing processor will deliver 3+ double-precision TFLOPS with 70+ cores and 400+ GB/s bandwidth with on-package high-bandwidth memory. Knights Landing primarily targets high-performance computing, but it also serves as a looking glass to glimpse future main-stream compute systems. Effectively utilizing its sheer compute power is however a challenge to software systems, especially those dealing with sparse data structures, increasingly used by emerging applications in domains such as high performance computing, machine learning, and data mining. This is because they typically have irregular parallelism and demand high memory bandwidth.A few case studies will be presented to showcase how sparse linear algebra algorithms can be efficiently mapped to such many-core and heterogeneous memory systems, followed by discussion on how the optimizations can be generalized in programming systems. First, our experience in optimizing sparse triangular solver demonstrates the need for inspector-executor optimization strategy, and decentralized synchronization mechanism instead of bulk synchronous parallelization with global barriers. This work led to #1 result in high performance conjugate gradient benchmark (HPCG), a new benchmark proposed as a complement to the famous high-performance LINPACK (HPL). Second, our optimization of HYPRE library demonstrates the need for code modernization that goes beyond the scope of a single function. This work resulted in up to 8x speedups of algebraic multigrid linear solver in HYPRE library, a popular solver originally developed by Lawrence Livermore National Lab and a prominent candidate for exa-scale solvers due to its asymptotic optimality.

Jongsoo Park is a research scientist in Parallel Computing Laboratory at Intel Corporation, where he has been studying emerging data-intensive high-performance computing applications and their implications on future processor architectures and programming systems. He received a Ph.D. from Stanford University (2011) and a B.S. from SNU (2005). During his Ph.D., he worked on a programming system for energy-efficient embedded processor. He is a recipient of the best paper award in Supercomputing conference 2012 for his work on low-communication FFT. His research on sparse iterative solver contributed to world's #1 HPCG benchmark results, and to algebraic multigrid package in Lawrence Livermore Lab's HYPRE library.