Communication Optimization in Iterative Numerical Algorithms: An Algorithm-Architecture Interaction

Trading communication with redundant computation can increase the silicon efficiency of common hardware accelerators like FPGA and GPU in accelerating sparse iterative numerical algorithms. While iterative numerical algorithms are extensively used in solving large-scale sparse linear system of equations and eigenvalue problems, they are challenging to accelerate as they spend most of their time in communication-bound operations, like sparse matrix-vector multiply (SpMV) and vector-vector operations. Communication is used in a general sense to mean moving the matrix and the vectors within the custom memory hierarchy of the FPGA and between processors in the GPU; the cost of which is much higher than performing the actual computation due to technological reasons. Additionally, the dependency between the operations hinders overlapping computation with communication. As a result, although GPU and FPGA are offering large peak floating-point performance, their sustained performance is nonetheless very low due to high communication costs leading to poor silicon efficiency. In this thesis, we provide a systematic study to minimize the communication cost thereby increase the silicon efficiency. For small-to-medium datasets, we exploit large on-chip memory of the FPGA to load the matrix only once and then use explicit blocking to perform all iterations at the communication cost of a single iteration. For large sparse datasets, it is now a well-known idea to unroll k iterations using a matrix powers kernel which replaces SpMV and two additional kernels, TSQR and BGS, which replace vector-vector operations. While this approach can provide a Θ(k) reduction in the communication cost, the extent of the unrolling depends on the growth in redundant computation, the underlying architecture and the memory model. In this work, we show how to select the unroll factor k in an architecture-agnostic manner to provide communication-computation tradeoff on FPGA and GPU. To this end, we exploit inverse-memory hierarchy of the GPUs to map matrix power kernel and present a new algorithm for the FPGAs which matches with their strength to reduce redundant computation to allow large k and hence higher speedups. We provide predictive models of the matrix powers kernel to understand the communication-computation tradeoff on GPU and FPGA. We highlight extremely low efficiency of the GPU in TSQR due to off-chip sharing of data across different building blocks and show how we can use on-chip memory of the FPGA to eliminate this off-chip access and hence achieve better efficiency. Finally, we demonstrate how to compose all the kernels by using a unified architecture and exploit on-chip memory of the FPGA to share data across these kernels. Using the Lanczos Iteration as a case study to solve symmetric extremal eigenvalue problem, we show that the efficiency of FPGAs can be increased from 1.8% to 38% for small- to-medium scale dense matrices whereas up to 7.8% for large-scale structured banded matrices. We show that although GPU shows better efficiency for certain kernels like the matrix powers kernel, the overall efficiency is even lower due to increase in communication cost while sharing data across different kernels through off-chip memory. As the Lanczos Iteration is at the heart of all modern iterative numerical algorithms, our results are applicable to a broad class of iterative numerical algorithms.Open Acces