|
Lower cost and manageability drive adoption of commodity-based HPC clusters
Not thinking seriously about HPC for your environment? If your organization needs greater performance, HPC can be a viable option, according to Bjorn Andersson, Sun's director of HPC.
No longer limited to the realm of nuclear physics, cosmological and environmental research, and other scientific and academic endeavors, HPC now offers mainstream industries an edge in real-time operations and complex calculations increasingly important to business performance, commented Andersson in a recent interview.
"HPC is being viewed as a competitive weapon that enables organizations to bring products to market faster with greater precision and to arrive at high quality decisions more quickly," he said.
What's Fueling the Growth of HPC
Lower cost, greater manageability, and an array of HPC hardware and software options are key drivers to broader HPC adoption. In the mid-1990s, for instance, a monolithic SMP, MPP, or single-processor supercomputer could cost $20 million and require a small army of specialized administrators.
Today, clusters of low-cost, commodity servers or blade modules dominate the HPC market. A low barrier to entry is helping to fuel HPC market growth that exceeded IDC's initial estimate of 20 percent a year, reaching $12 billion in 2008 for servers alone — and outpacing the general IT market four to one.
And with cluster management software, organizations have a way to administer HPC installations ranging from a handful to hundreds or thousands of nodes from a single console, according to Andersson. This reduces administrative overhead and enables greater focus on the core business challenges that HPC is geared to solve with extreme performance clocked in teraflops, or trillions of floating point operations per second.
"Sun's focus is to make HPC affordable and usable by the mainstream, while delivering the always-higher performance that traditional HPC users demand," Andersson said.
How to Get Started with HPC
Is HPC a viable option for your organization? Sun offers several easy ways to "kick the tires" and assess HPC benefits, as well as an end-to-end platform of servers, blades, storage systems, and software that supplies choice and flexibility at every step of the way.
Sun HPC Solution Centers:
Located in Oregon and Scotland — and accessible from smaller centers around the world — these multiple teraflop x64 clusters offer customers proofs of concept on how HPC can answer business demands.
Sun Customer Ready HPC Clusters:
The program lets you configure systems online and purchase factory-integrated, ready-to-deploy HPC clusters complete with blades or servers, networking, and OS and grid software. Sun HPC Quick Start Services offers expertise in architecting and implementing HPC systems.
Choosing the Right Platform and Processors
One question is whether to use rackmount servers or blade modules to build an HPC cluster. Blades are generally easier to implement and manage, with preintegrated networking and simplified power supplies that can speed deployment and ease maintenance, Andersson said. Servers can make sense for low-end or prototype installations by avoiding the slightly higher up-front chassis costs of blades, though managing multiple hardware nodes can introduce complexity and hidden costs.
"There's a little higher threshold getting into blades because of the chassis, but once you pass that threshold it's much more efficient and easier to deal with — and you're protecting your investment on a different level," Andersson said. "For instance, if two years down the road you want to upgrade your CPUs, with a blade approach it's easier and your power supply, networking, and everything else remains intact."
Another question is processor type. Sun offers customers a choice of the latest Intel Xeon and AMD Opteron chips in its x64 server and blade module lineup; for instance, quad-core Intel Xeon processors were introduced in June into the new Sun Blade X6450 module, delivering up to 7.37 teraflops in a single rack, Andersson noted.
Options for Multithreaded HPC Applications
Sun's chip multithreading (CMT) UltraSPARC T2 CPU is also an option for organizations running multithreaded HPC applications and with a need for the higher throughput available with CMT technology. Unlike the first-generation UltraSPARC T1 processor, geared for high Web scale throughput, UltraSPARC T2 introduced the floating point support that HPC requires. In addition, blades using different CPUs (including CMT) can be mixed in a cluster and Sun Grid Engine can automatically deploy applications where they run best.
"For certain types of applications, the UltraSPARC T2 'Niagara' technology works extremely well," Andersson said. "We're seeing a lot of interest in UltraSPARC T2 because it's really well suited for applications that scale well on multi-core, multithreaded systems."
For instance, Andersson noted, the High Performance Computing Virtual Laboratory (HPCVL) in Canada this year implemented a 78-node HPC cluster based on Sun SPARC Enterprise T5140 servers with UltraSPARC II processors.
"The CMT technology built into the Sun SPARC Enterprise T5140 server has enabled us to help some of our 600+ researchers begin to use the new multithreaded, multi-core technologies and prepare to scale rapidly to support the massive workload on the servers," said Dr. Ken Edgecombe, HPCVL executive director. "We are seeing scaling even better than expected as we continue to see application demand increase."
"Sun's focus is to make HPC affordable and usable by the mainstream, while delivering the always-higher performance that traditional HPC users demand."
Growing Demands for HPC Storage
As HPC clusters grow, so too does the need for HPC storage. "If you look at the IDC numbers, the storage side of HPC is growing even faster than the computer side, and the computer side is growing on a double-digit curve," Andersson said.
At the Texas Advanced Computing Center (TACC), 72 Sun Fire X4500 server/storage hybrids combine for a massive 1.7 petabytes of raw storage capacity, with a long-term retention and archiving solution scalable to 3.1 petabytes. This supplies the storage infrastructure for TACC's "Ranger" supercomputer, based on nearly 4000 Sun Blade modules with more than 500 teraflops of peak compute power. It's ranked #4 on the Top 500 Supercomputers List.
Also prominent in the HPC storage space is the Lustre file system, which Sun acquired in September 2007. As Andersson noted, six of the top 10 sites on the Top 500 Supercomputers List run Lustre. The parallel, open source Lustre software enables aggregation of tens of thousands of cluster nodes and petabytes of storage into a single object-oriented system, and can achieve more than 90 percent raw bandwidth I/O.
"With Lustre, you can have a single file system with a unified name space across multiple servers and storage systems, and that simplifies manageability from the user and application standpoint," Andersson said. "And because Lustre is a parallel system, you can get much higher bandwidth and performance." One Lustre user is Chevron, which uses it to handle exponential growth in data from oil exploration.
In Sun's HPC Pipeline
With heightened interest in HPC, Sun is gearing up to highlight additional advancements to its HPC solutions set at the 20th annual SuperComputing 2008 (SC08) Conference, November 15-21 in Austin, Texas. According to Andersson, Sun is continuing to focus efforts on usability and manageability for mainstreams users, while pushing the petascale frontier of HPC performance.
"It always comes back to addressing core business issues and turning HPC into a competitive advantage," Andersson said. "To get there, customers need HPC solutions with low barrier to entry and rapid payback. That's why our focus is on flexible and manageable HPC solutions that can extend from a small cluster in a single rack to some of the largest supercomputers in the world today."
SUN'S HPC DEVELOPMENT AND APPROACH
The manageability, performance, and flexibility that define Sun's approach to HPC were reflected in enhancements to Sun's HPC portfolio unveiled in June at the International Supercomputing Conference in Dresden, Germany, Andersson noted. Sun's approach focuses on three key areas:
1) HPC Manageability and Development
Return on investment from an HPC initiative can be diminished if installation, management, and application development consume inordinate resources. Flexibility to interoperate with heterogeneous systems is important to extend HPC's reach.
Single-console administration of multi-node clusters, policy-based workload management, and dynamic workload provisioning can all boost efficiency and reduce cost and complexity. On the development side, HPC-optimized tools can help accelerate creation and deployment of HPC applications. Learn about:
2) HPC Performance
HPC is fundamentally about ever-greater performance. As eWeek magazine noted, the Sun Constellation System is pushing the petaflop boundary with such installations as the TACC Ranger cluster. And Sun's system won the Supercomputing Online Editor's Choice Award for 2007 Product of the Year. Learn what's new with:
3) HPC Storage Flexibility and Capacity
Even a low-end HPC cluster will process mammoth amounts of data. Storing and archiving that information is critical to HPC value from both an IT and business perspective, yet balancing high-performance data access with economical long-term data storage can be a challenge.
Sun's approach focuses on flexible, high-capacity Open HPC storage systems tailored to customer needs. An ideal solution will utilize high-performance disk storage for regularly accessed data, bulk disk storage for lower performance needs, and energy-saving tape for long-term storage. Learn how:
|
|
