Hardware-Assisted Context Management for Accelerator Virtualization: A Case Study with RSA

The advantages of virtualization, including the ability to migrate, schedule, and manage software processes, continues to drive the demand for hardware and software support. However, the packaging of software state required by virtualization is in direct conflict with the trend toward accelerator-rich architectures where state is distributed between the processor and a set of heterogeneous devices --- a problem that is particularly acute in the mobile SoC market. Virtualizing such systems requires that the VMM explicitly manage the internal state of all of the accelerators over which a process's computation may be spread. Public-key crypto engines are particularly problematic because of both the sensitivity of the information that they carry and the long compute times required to complete a single task.

In this paper we examine a set of hardware design approaches to public-key crypto accelerator virtualization and study the trade-off between sharing granularity and management overhead in time and space. Based on observations made during the design of several such systems, we propose a hybrid local-remote scheduling approach that promotes more intelligent decisions during hardware context switches and enables quick and safe state packaging. We find that performance can vary significantly among the examined approaches, and that our new design, with explicit accelerator support for state management and a modicum of scheduling flexibility, can allow highly contended resources to be efficiently shared with only moderate gains in area and power consumption.