Academy & Industry Research Collaboration Center (AIRCC)

Volume 10, Number 11, September 2020

VirtFun: Function Offload Methodology to Virtualized Environment

  Authors

Carlos A Petry and Rodolfo J. de Azevedo, University of Campinas, Brazil

  Abstract

The use of virtual machines (VM) has become popular with substantial growth for both personal and commercial use, especially supported by the progress of hardware and software virtualization technologies. There are several reasons for this adoption like: cost, customization, scalability and flexibility. Distinct domains of application, such as scientific, financial and industrial, spanning from embedded to cloud systems, taken advantage of this kind of machines to meet processing computational demands. However, there are setbacks: hardware handling, resources use, performance and management. This growth demands an effective support by the underlying virtualization infrastructure, which directly affects the hosts’ capacity in datacenters and cloud environment that support them. It is evident that the host native processing performs better than VMs, especially when using accelerator devices, where the common solution is to assign each device to a specific VM, instead of sharing it among multiples VMs. Beyond performance issues inside the host, we need to consider the VM performance when using accelerator devices. In this context, it is necessary to provide efficient mechanisms to manage and run VMs which can take advantages of high-performance devices, like FPGAs or even from software resources on the host. To assist this challenge, this paper proposes a methodology to improve communication performance of applications running on the VMs, VirtFun. To do so, we developed a framework able to offload pieces of application's code (vFunction) to host by means of secure data sharing between the application and device. The results achieved in our experiments demonstrated significant acceleration capacity for the guest application vFunction. The speedup reached 340% compared to conventional network execution, reaching maximum slowdown of 2.8% in the worst case and near to 0% in the best case considering the native execution.

  Keywords

Virtualization, performance, virtual machine, shared-memory.