|Ph.D Student||Amit Nadav|
|Subject||Alleviating Virtualization Bottlenecks|
|Department||Department of Computer Science||Supervisors||Professor Assaf Schuster|
|Professor Dan Tsafrir|
|Full Thesis text|
Hardware virtualization has long been studied, but has only recently become popular, after being introduced to commodity servers. Despite the ongoing research and the developing hardware support, virtual machines incur degraded performance in a wide variety of cases, especially when an unmodified virtual machine operating system is used. One of the major causes of this degraded performance is the lack of physical hardware transparency in virtual machines, since the hypervisor---their controlling software-layer---usually exposes hardware abstractions instead of the physical hardware. While such abstractions are often required to multiplex hardware in virtualization environments, they introduce inefficiencies.
In our work we investigate a wide variety of scenarios in which the lack of transparency incurs substantial performance overheads: I/O memory management unit (IOMMU) emulation, interrupts multiplexing by the hypervisor and memory over-provisioning. For each of these scenarios we suggest novel methods to increase transparency without the virtual machine's cooperation, and thereby improve performance without modifying its operating system and without access to its source code. Accordingly, the methods we propose apply to proprietary operating systems as well and ease the porting of virtual machines from one hypervisor to another.
First, we show that virtual machine performance with IOMMU emulation, which enhances security, can improve by up to 200% using a novel sidecore emulation approach---performing device emulation by another core instead of the hypervisor. Second, we present a secure and efficient method for selective delivery of interrupts to virtual machines, improving performance by up to 60% for I/O intensive workloads. Last, we introduce VSWAPPER, an efficient uncooperative swapping extension that enhances VM performance when memory is overcommitted by up to an order of magnitude.