Closing the Benchmark Gap for Tiered Memory
Rajath Shashidhara, Simon Peter, Scott Hare, Kimberly Keeton · 2025
The increasing adoption of tiered memory in modern data centers requires reliable benchmarks to evaluate their performance. Unfortunately, existing benchmarks often fail to reflect the complexity and variability of memory access patterns of production workloads, leading to design gaps. This paper systematically explores the divergence between benchmarks and real-world memory access patterns. We propose statistical models, fleet-wide telemetry, and workload similarity metrics and discuss how they may be combined into a datadriven framework to generate representative benchmarks with high fidelity and wide coverage of production behaviors. We outline challenges for this approach and a path towards automated synthesis of production-representative benchmarks.