Tiered Memory Management Beyond Hotness
summary
Introduces amortized offcore latency (AOL), a metric that divides off-chip latency by observed MLP to predict performance impact; builds an offline profiler-driven allocator SOAR that ranks objects by AOL-weighted contribution (uses PBS traces and CPU counters) and an online migration regulator ALTO that gates promotions with two AOL thresholds. Evaluated on remote-NUMA and CSL hardware against four state-of-the-art tiering systems, achieves up to 12.4× speedup and never falls behind by more than 3%.
AI-written summary · about