SC17 Denver, CO

P54: Investigating Hardware Offloading for Reed-Solomon Encoding

Authors: John W. Dermer (Los Alamos National Laboratory), Gustavo De Leon (Los Alamos National Laboratory; University of California, Berkeley), Tyler S. Rau (Los Alamos National Laboratory)

Abstract: Reed-Solomon (RS) encoding is a storage scheme which offers better scalability, but requires heavier computation, compared to other models. This presents a problem as it requires users to purchase brawnier CPUs than would be otherwise necessitated. However, Mellanox’s ConnectX-4 Infiniband cards have the capability to perform RS encoding on the HCA hardware; removing the need for powerful CPUs to calculate it. We investigated the performance, measured in throughput, between these cards and Intel’s ISA-Library, with regard to various block sizes and concurrency. We conclude that the MLX cards encoded faster and more consistently than ISA-L. Furthermore, the ConnectX-5 cards support the Galois Field (GF) 2^8, this grants compatibility with data encoded by any system using GF(2^8) or less, including ISA-L. These cards enable users to substantially increase encoding and decoding throughput by using more cards; additionally enabling the use of less powerful CPUs to achieve similar performance. LA-UR-17-26333
Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf

Poster Index