High Performance Virtual Machines Project
in the Concurrent Systems Architecture
Group
We are building software technology for scalable clusters which leveraging
commodity platforms, hardware (PC's, servers, and workstations) and software
(Windows NT, Linux). While many commercial efforts in this area focus on
transaction processing systems, we are focused on scalable performance
for a broader range of applications including emerging national information
infrastructure applications (scalable
web serving,interactive visualization/data
exploration, information
serving, data
mining, etc.) as well as traditional high
performance computing applications (HPC). Synthesizing usable ensembles
from distinct hardware systems, and independent operating system kernels
requires high bandwidth, low latency communication and efficient coordination
within clusters and high performance networking outside the cluster. To
meet these demands, we are researching a variety of technologies:
-
High speed, lightweight communication layers which provide low-latency
and high bandwidth (101MB/s = 808Mbps delivered application bandwidth,
10 microsecond latencies)
-
Cluster resource coordination (scheduling or processors, memory, and input/output)
to improve cluster performance (Dynamic Coscheduling)
-
Fastpaths and parallel protocol processing to improve networking performance
(higher bandwidth, higher packet rates, and larger numbers of connections)
-
Enhanced protocols to support continuous media and QoS in both lightweight
and traditional protocols
-
New technologies which support applications with dynamic communication
structures in dynamic environments
The HPVM software provides basic turnkey computing for HPC clusters. It
includes high speed communication layers, tools for
coordinated resource management, and a Java-based front
end for cluster access and management.
HPVM Application Programming Interfaces (API's):
HPVM Testbeds
We are using several testbeds and interconnects. These allow us to explore
a variety of hardware and operating system environments.
-
the CSAG
HPVM grows to 128 Procs...
-
HPVM
III A 256-processor 300 Mhz Pentium II Cluster with Myrinet and 100Mbit
ethernet); Windows NT
-
HPVM
II A 64-processor 300 Mhz Pentium II Cluster with Myrinet, Servernet,
switched 100Mbit ethernet); Windows NT and Linux
-
HPVM
I A 30-processor Pentium Pro Cluster with Myrinet and 100Mbit Ethernet);
Windows NT and Linux
CSAG also maintains a page documenting Resources
for High Performance Computing in Windows NT.
HPVM Questions
HPVM Bug Reports
Back to CSAG home page
Last updated 8 January 1998
webmaster