Reliability-aware Garbage Collection for Hybrid HBM-DRAM Memories

Emerging workloads in cloud and data center infrastructures demand high main memory bandwidth capacity. Unfortunately, DRAM alone is unable to satisfy contemporary demands. High-bandwidth (HBM) uses 3D die-stacking deliver 4–8× higher bandwidth. HBM has two drawbacks: (1) capacity low, (2) soft error rate high. Hybrid combines promise low fault rates, bandwidth, Prior OS approaches manage by mapping pages versus based on hotness (access frequency) risk (susceptibility errors). these operate at a coarse-grained page granularity, frequent migrations hurt performance. This article proposes new class of reliability-aware garbage collectors for hybrid HBM-DRAM systems that place hot low-risk objects the rest DRAM. Our analysis nine real-world Java shows that: newly allocated nursery are frequently written, making them both low-risk, small fraction mature (3) allocation site good predictor risk. We propose RiskRelief, novel collector prediction HBM. Allocation sites profiled offline RiskRelief heuristics classify as The proposed expose Pareto-optimal trade-offs between (SER) execution time. improves SER 9× compared an HBM-Only system while same time improving performance 29% DRAM-Only system. Compared state-of-the-art approach placement, eliminates all migration overheads, which substantially delivering similar SER. Reliability-aware collection opens up opportunity emerging memories fine granularity requiring no extra hardware support leaving programming model unchanged.