Coordinated and efficient huge page management with ingens

Abstract

Modern computing is hungry for RAM, with today's enormous capacities eagerly consumed by diverse workloads. Hardware address translation overheads have grown with memory capacity, motivating hardware manufacturers to provide TLBs with thousands of entries for large page sizes (called huge pages). Operating systems and hypervisors support huge pages with a hodge-podge of best-effort algorithms and spot fixes that made sense for architectures with limited huge page support, but the time has come for a more fundamental redesign. Ingens is a framework for huge page support that relies on a handful of basic primitives to provide transparent huge page support in a principled, coordinated way. By managing contiguity as a first-class resource and by tracking utilization and access frequency of memory pages, Ingens is able to eliminate a number of fairness and performance pathologies that plague current systems. Experiments with our prototype demonstrate fairness improvements, performance improvements (up to 18%), taillatency reduction (up to 41%), and reduction of memory bloat from 69% to less than 1% for important applications like Web services (e.g., the Cloudstone benchmark) and the Redis key-value store.