V4VSockets: low-overhead intra-node communication in Xen.

Nowadays, the cloud computing paradigm has given rise to new and unconventional application deployments on elastic compute infrastructures. For instance, IaaS providers are willing to support a diverse set of computing workloads, ranging from service-oriented deployments to HPC applications. As a result, the underlying systems software has to be generic enough to support lightweight, efficient execution for a wide range of applications. In this work, we examine communication methods within a single VM container that ease data exchange between co-located VMs without sacrificing upper-layer API compatibility. We present V4VSockets, a generic, socket-compliant framework for intra-node communication in the Xen hypervisor. The transport layer is based on V4V, a simple hypercall-based mechanism to transfer data. Our framework resides within the hypervisor, providing a dispatch logic to communication, contrary to the common Xen concept of decoupling the control and data plane using a privileged VM. V4VSockets improves intra-node data exchange in terms of both latency and throughput by a factor of 4.5. To demonstrate the applicability of V4VSockets, we spawn a VM with a GPU device assigned to it and deploy a remote GPU acceleration benchmark on co-located VMs. Preliminary results show that V4VSockets boosts the transfer throughput by a factor of 7 (at best) while adding an overhead of 15% compared to native execution.