Providing a cloud network infrastructure on a supercomputer

Abstract

Supercomputers and clouds both strive to make a large number of computing cores available for computation. More recently, similar objectives such as low-power, manageability at scale, and low cost of ownership are driving a more converged hardware and software. Challenges remain, however, of which one is that current cloud infrastructure does not yield the performance sought by many scientific applications. A source of the performance loss comes from virtualization and virtualization of the network in particular. This paper provides an introduction and analysis of a hybrid supercomputer software infrastructure, which allows direct hardware access to the communication hardware for the necessary components while providing the standard elastic cloud infrastructure for other components.

Light-weight remote communication for high-performance cloud networks

“…Wang and colleagues have extended Hadoop to take advantage of InfiniBand [3]. Similarly in spirit, works by Jose et al [4] and by the authors themselves [5] have investigated how a distributed memory cache can exploit HPC interconnects such as InfiniBand or those of the Blue Gene/P supercomputer for better efficiency. LibRIPC strives to generalize such efforts both with respect to the cloud workloads running atop and the hardware architectures running underneath.…”

Evaluating Load Generation in Virtualized Environments for Software Performance Testing

“…When applied 100% CPU, they observed a 26.5% CPU usage for Dom-0, which is a similar result we got in our experiments. There are also other works that try to minimize the I/O overhead imposed by virtualization, for instance, Appavoo et al [18] presented a mechanism to improve performance of network access in virtualized machines, and Wei et al [19] investigated the use of dedicated Xen domains to handle I/O operations. Liu and Abali [20] proposed the Virtualization Polling Engine (VPE) that uses dedicated CPU cores to help with the virtualization of I/O devices by using an eventdriven execution model with dedicated polling threads.…”

A light-weight virtual machine monitor for Blue Gene/P

“…We have verified our implementation with Kittyhawk Linux, a BG/P version of the Linux Kernel [4] with support for BG/P's hardware devices. It also provides an overlay that maps standard Linux Ethernet communication onto Blue Gene's high-speed collective and torus interconnects [3]. Our VMM allows one or more instances of Kittyhawk Linux to run in a VM.…”

“…Figure 7 shows the results. Our virtualization layer poses significant overhead to the Ethernet network performance (which is already off the actual performance that the torus and collective hardware can deliver [3]). We are confident, however, that optimizations such as para-virtual device support can render virtualization substantially more efficient.…”