Towards low-latency and reliable RAID caching with ERcache

Parity-based Redundant Array of Independent Disks (RAID), specifically RAID 5, use distributed parity to offer device-level fault tolerance, thereby increasing the reliability of storage systems. It also provides balanced trade-off between throughput and cost. However, RAID 5 suffers from a significant performance bottleneck known as the write penalty, where partial stripe writes necessitate extra I/Os for parity updating, which can reduce system endurance and reliability. To mitigate the extra I/Os caused by the write penalty, traditional stripe-based LRU caches attempt to delay flushes to compact requests into full stripes. Despite this, these caches exhibit severe write amplification and low hit ratios, which remains inefficient. Our analysis reveals two primary factors causing these inefficiencies: the repeated partial stripe evictions and the long time staying of cold data within full stripes. To address these challenges, we propose ERcache, a novel RAID cache scheme designed to minimize write volume and improve hit ratios. ERcache selectively retains partial stripes slated for eviction by predicting the return time based on key access features, thereby avoiding redundant I/Os. In addition, ERcache identifies the hotness of data blocks to reorganize the full stripes, allowing the system to reclaim cache space without triggering the write penalty. Our experiments, encompassing 12 workloads across multiple cache sizes, demonstrates that ERcache reduces average I/O latency by more than 38.3% compared to state-of-the-art academic RAID cache techniques and mature commercial hardware controllers. Meanwhile, ERcache reduces average write volume by 13.7% compared to traditional RAID cache schemes, enhancing system reliability.