Fermilab Computing Division

Architecting a Symbiotic Virtual Machine Monitor for Scalable High Performance Computing

Simple document list
(1 extra document)

Full Title: Architecting a Symbiotic Virtual Machine Monitor for Scalable High Performance Computing
Date & Time: 26 Jan 2010 at 11:00
Event Location: FCC2A/2B
Event Moderator(s):
Event Info: Title: Architecting a Symbiotic Virtual Machine Monitor for Scalable High
Performance Computing

Speaker: Jack Lange

Speaker Bio:
Jack Lange is a CS Ph.D. candidate in the Department of Electrical
Engineering and Computer Science at Northwestern University. He
earned his B.S. degrees in Computer Science and Computer Engineering,
and his M.S. degree in Computer Science from Northwestern University.
Jack's research interests include virtualization, high performance
computing (HPC), operating systems (OS), networking, and empathic
systems. He has published 13 papers on these topics. He is the lead
graduate researcher on the Palacios Virtual Machine Monitor (VMM), an
OS-independent VMM designed for research and use in high performance
computing as well as operating systems and architecture
research. Palacios is the focus of a large collaboration between
Northwestern University, Sandia National Laboratories and the
University of New Mexico. In 2007, Jack was one of three recipients
nationwide of the Symantec Graduate Fellowship. Jack has been a
program committee member for 7 workshops and conferences, and is
currently the student activities co-chair for the HPDC (High
Performance and Distributed Computing) conference for 2010 in Chicago.

While virtual machines (VMs) have become commonplace in datacenter and
cloud computing environments, they have yet to make inroads into high
performance computing (HPC). Despite this, there is considerable
interest in bringing the well established benefits of virtualization
to that domain. However, there are many concerns about parallel
application performance and scalability. Currently, there are
currently no virtual machine monitor (VMM) architectures designed with
HPC in mind.

In response, we have developed the Palacios VMM, an OS-independent VMM
designed specifically for research and use in HPC, as well as for
systems and architecture research. Palacios is part of a collaborative
research effort between Northwestern University, the University of New
Mexico, and Sandia National Laboratories. As of December, Palacios has
successfully been tested at scales exceeding 4096 nodes on the
RedStorm Cray XT supercomputer at Sandia, the largest scale
performance study of virtualized high performance computing ever
conducted. Overheads were extremely low, and Palacios-based
virtualization had minimal (less than 5%) impact on scalability.
These results demonstrate the practicality of virtualization in high
performance computing. Palacios is an open source project, and has
been downloaded over 1,000 times as of July 2009. This talk will focus
on the Palacios VMM and its use in HPC environments.

I will also describe symbiotic virtualization and its integration
inside the Palacios VMM. Symbiotic virtualization is a new approach to
the interface between the guest OS and the VMM in which the OS makes a
high-level semantically rich interface available to a VMM. The VMM can
use the interface to asynchronously or synchronously acquire
information from the guest. This interface is not required for correct
functionality, nor does it prevent an OS from running on physical
hardware. When the symbiotic interface is present, a VMM can leverage
the information it provides to enable optimizations and services that
are difficult or impossible to implement given current VMM/guest
structures. I will show an example optimization, SwapBypass, that is
enabled by symbiotic virtualization, and describe my current and
future efforts in applying symbiotic virtualization to HPC and other

No talks in agenda

Other documents for this event

CS-doc-# Title Author(s) Topic(s) Last Updated
3727-v1 presentation slides Jack Lange High Performance Computing
12 Mar 2010

DocDB Home ]  [ Search ] [ Authors ] [ Events ] [ Topics ]

DocDB Version 8.8.9, contact Document Database Administrators
Execution time: 1 wallclock secs ( 0.20 usr + 0.04 sys = 0.24 CPU)