Hzmem: New Huge Page Allocator with Main Memory Compression

Abstract

Today, applications that require large memory footprint prevail in cloud computing fields from both industry and academia. They impose great stress on the memory management of operating system, spend quite a substantial proportion of time dealing with TLB misses and excessively reduce consolidation ratio in term of server consolidation and virtualization. There are two methods to address these problems: main memory compression and large page support. However, to the best of our knowledge, there is no existing practical research on combination of these two methods since the combination in commodity operating system requires lots of low-level design modifications.

We propose a new memory management framework that is decoupled and flexible for easy developments and is able to run simultaneously with the original memory management. On top of the new framework, we implement Hzmem, a completely new large page memory management redesign with features of main memory compression to address the aforementioned problems once for all but requires only minor modifications to the other subsystems of the underlying operating system. Our method achieves competitive performance compared with native large page support, increases effective memory size and impacts little on other subsystems of operating system.