Research

This work asks whether memory reclamation has memory of its own. When a kernel preferentially evicts cold pages, the remaining resident set becomes hotter; an equal reclaim action can therefore impose a different future cost until evicted pages refault. The current theory models that recovery from the victim-conditioned access-rate distribution and represents each tenant’s accumulated history as reclaim debt.

The research is now moving from theory to measurement. The next tests estimate page-access-rate distributions, preregister recovery-curve predictions, and compare randomized reclaim against a uniform-access negative control. The access-rate estimator, cold-frontier controller, history-aware water-filling policy, and randomized assay are not yet implemented; existing measurements characterize the experimental substrate only.

This work focused on generating hardware-efficient code for datacenter workloads. I built compiler passes for SIMD vectorization, loop tiling, and cache-aware scheduling, then extended an LLVM/MLIR pipeline with fusion, common subexpression elimination, and strength reduction to reduce memory traffic.

The same pipeline used hardware-aware cost models for cache and NUMA effects, instruction throughput, and memory bandwidth to guide auto-tuning across CPU and GPU kernels. On representative benchmarks, the combined work improved end-to-end latency by 32.478% and throughput by 15.821%.