Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Learn about advanced tiered memory management techniques that move beyond traditional hotness-based approaches in this 16-minute conference talk from OSDI '25. Discover how researchers from Virginia Tech introduce amortized offcore latency (AOL), a novel metric that more accurately captures the true performance impact of memory accesses by considering both memory access latency and memory-level parallelism (MLP). Explore two innovative tiering mechanisms: SOAR, a profile-guided allocation policy that strategically places objects based on their actual performance contribution rather than just access frequency, and ALTO, a lightweight page migration regulation policy designed to eliminate unnecessary data migrations. Understand how these approaches address the limitations of hotness-based policies, where frequently accessed data may not always be performance-critical, and examine performance results showing improvements of up to 12.4× over four state-of-the-art tiering designs across diverse workloads, with minimal underperformance in edge cases.
