High-Performance Cluster for Biomedical Research Using 10 Gigabit Ethernet iWARP Fabric
A High-Performance Cluster for Biomedical
Research Using 10 Gigabit Ethernet iWARP Fabric
A large research institute has achieved performance of nearly 36 TeraFLOPS at greater than 84 percent efciency using
the HPL benchmark on a cluster of 4,032 cores. iWARP enabled these results using 10 gigabit Ethernet (10GbE) by
reducing the overhead associated with kernel-to-user context switches, intermediate buffer copies, and TCP/IP processing.
A large research facility has
achieved excellent performance
and near-linear scalability using
iWARP and NetEffect™ 10 gigabit
Ethernet Server Cluster Adapters
on a cluster of 4,032 cores, as
measured using the HPC LINPACK
benchmark. This result represents
a relatively low-cost approach to
processing very large technical
workloads using commercial off-
the-shelf network hardware.
EXECUTIVE SUMMARY
Using ubiquitous, standards-based Ethernet technology, iWARP (Internet Wide Area
RDMA Protocol) enables low-latency network connectivity suitable for high-performance
clusters. A key advantage of iWARP networking is its compatibility with existing network
infrastructure, management solutions, and solution stacks.
This paper demonstrates the viability of cluster computing based on iWARP to achieve
very high performance using 10GbE. It begins with a description of the architecture
of a cluster based on iWARP connectivity before moving to a brief overview of iWARP
technology. The paper concludes by reporting performance achieved using that cluster
and observations about the value of iWARP to future work in this area.
Architecture of an iWARP
Cluster for Medical Research
To support large-scale workloads in a
range of areas critical to its research,
including bioinformatics, image analysis,
and sequencing, a research institution has
built a large (4,032 cores) cluster using
iWARP. For the compute nodes, they chose
two-way Dell PowerEdge* R610 servers
based on Intel® Xeon® processors x5550
running at 2.66 GHz with 24 GB RAM and a
single 80 GB SATA hard drive in each server.
For RDMA (remote direct memory access)
network connectivity, the design uses
NetEffect™ 10GbE Server Cluster Adapters.
June 2010
Version 1.1
Tom Stachura
Intel LAN Access Division
Brian Yoshinaka
Intel LAN Access Division
WHITE PAPER
Internet Wide Area RDMA Protocol (iWARP)
NetEffect™ 10 gigabit Ethernet
Server Cluster Adapters
Technical and High-Performance Computing