Kunpeng Instruction-based ISA-L Optimization

Overview

The Intelligent Storage Acceleration Library (Intel® ISA-L) is widely used in distributed storage system software. This feature accelerates the CRC32 and EC algorithm APIs based on Kunpeng vectorized instructions.

Technical Principles

• CRC32 algorithm optimization

  Optimization 1: The 3-way CRC32 instruction parallelism of open source ISA-L can be expanded to 6-way parallelism on the new Kunpeng 920 processor model. This optimization improves the computing performance under cache full-hit conditions. Furthermore, prefetch optimization tailored for small data blocks are integrated to improve the performance during cache miss occurrences.

• EC algorithm optimization

  Optimization 1: The first parity block of EC calculation is calculated by using the XOR operation instead of table lookup.

  Optimization 2: Prefetch offset optimization is added to improve the cache hit ratio.

Expected Results

The optimized CRC32 and EC (EC 10+1, 4+2, and 8+3) algorithms deliver 10% and 20% performance gains, respectively, over open source ISA-L 2.31.