Reproducing end-to-end performance via MP-based artificial fragmentation

Establish whether reproducing physical memory fragmentation patterns recorded from /proc/kpageflags using Anduril and Markov Process profiles suffices to reproduce system-wide fragmentation effects and end-to-end workload performance metrics, including runtime, throughput, tail latency, peak memory usage, and huge page usage.

Background

The authors propose a two-step validation: (1) reproduce fragmentation patterns from /proc/kpageflags with Anduril, and (2) show that reproducing those patterns reproduces system-wide performance effects. They report that step (1) often succeeds, but step (2) fails to match end-to-end performance observed under other fragmentation methodologies.

This failure leaves open whether the MP-based reproduction approach can be refined to achieve performance fidelity, and under what conditions or modeling enhancements (e.g., temporal information, partition awareness) it might succeed.