Remote Memory Swapping for Virtual Machines in Commercial Infrastructure-as-a-Service

Abstract

Memory swapping is an essential feature offered by operating systems (OS) to prevent applications from crashing due to the system running out of main memory. Traditional swapping works by using a part of the storage disk as backup memory. Unfortunately, storage disk speeds are orders of magnitude slower than main memory. This implies that as useful as swapping is, it causes expected performance degradation of the application. A more recent swapping mechanism in clusters is remote swapping, which uses one host’s main memory as swap space over the network for another host. For faster networks, remote swapping could outperform disk swapping. However, for users leasing Virtual Machines (VMs) in existing commercial Infrastructure-as-a-Service (IaaS) providers, choosing a swapping mechanism for their VMs is non-trivial. They are subjected to additional constraints and considerations such as lack of access to the underlying hypervisor, OS, and hardware. On the other hand, network speeds within clouds have evolved to be faster than storage device speeds, making remote swap appealing. This paper explores implementing remote swapping for leased VMs within IaaS providers. A setup is proposed that implements remote swapping within a cluster of VMs, abiding by the constraints of IaaS. The setup does not need any modification of the hypervisor, or any special hardware, making it suitable for users currently in commercial IaaS clouds. Through extensive experimental evaluations on Amazon Web Services’ ec2 instances and Google Cloud’s Compute Engine, the paper demonstrates the performance benefits of using remote memory versus the standard storage disks as swap space.