High Performance Erasure Coding - the Story so Far
summary
Work to make erasure coding perform near replica‑3 claims concrete: a Python offline consistency checker and C++ io exerciser plus an online checker validate 4+2 pools; read path gains minimum required read (and partial stripe read) and write path implements minimum required write to read only equivalent blocks, cutting IO and network amplification. A dual buffer change (payload vs raw) removes large memmove stalls (~8% CPU in one profile) and reduced metadata updates. Prototype branch shows notably higher IOPS for 4K/8K random reads and writes versus Squid.
AI-written summary · about