SYSTEMS

Jonathan Follows, Team Leader of the Distributed Computing Group (DisCo) at Daresbury Laboratory said “we look forward to working with all of the partners in this program and thank Streamline Computing for implementing this system within a very tight timescale to meet STFC requirements”

Daresbury Laboratory is planning for the Hartree Centre which will address "grand challenge" science delivered through high performance computing, to be implemented in 2011. One significant part of this is expected to be delivered using GPU computing and this latest cluster represents a collaborative programme between Daresbury, Streamline and NVIDIA to explore the possibilities of this technology.

Technical specialists from NVIDIA and Streamline Computing will work together to ensure that the cluster runs to maximum efficiency but some early research from Streamline Computing has already demonstrated 10 fold speedups in CUBLAS and CUFFT libraries. Further collaborations may also be possible with software tool and compiler companies to test new tools for GPU computing.

”We very much welcome the opportunity to continue to work with the excellent technical team at STFC Daresbury Laboratory and together enhance the skills and experience of Streamline engineers allowing us to continue to deliver best of breed high performance computing clusters making the best use of commodity technology” said Dr John Taylor, CTO of Streamline Computing. “The combination of Nehalem and Tesla are reaching new plateaux in commodity performance and the solution at Daresbury will provide a major research platform for scientific computing”

There are some important advances in technology and systems design incorporated within the new cluster, including:

• Intel Nehalem based nodes with higher I/O and memory bandwidth and QuickPath Interconnect
• Dual socket nodes now supporting multiple NVIDIA Tesla cards via a dual PCI-E (16) Gen 2 provided by Supermicro
• Each resilient node supports a complete Tesla S1070 1U server chassis with 4 X 1.29GHz Tesla cards
• Streamline Computing solution containing a CUDA ready software stack installed on the system complete with pre-installed tools and libraries so the research team at Daresbury can utilise the cluster from day 1 after installation
• The whole solution is managed through a concurrentCOMMAND cluster management appliance from Concurrent Thinking Ltd

The cluster has been designed to show how effective GPU Computing can be for many classes of application but linked with Nehalem they also allow Daresbury to ensure the minimum of bottlenecks when connecting GPUs through Infiniband in each node to high performance storage systems. This additional cluster enhances a facility installed in the previous year which utilised Harpertown and Seaburg based nodes with the then fastest IB solution with DDR and PCI-E Gen 2 connect X cards from Mellanox.

Streamline Computing are particularly grateful to Supermicro for early release server products which demonstrate Nehalem performance in practical and reliable server solutions. General Manager of Supermicro in Europe, Dev Tyagi, said "Supermicro has an excellent track record for quickly implementing new technologies from companies like Intel to provide early adopters and pre-release customers timely access to next-generation server products. The company's ability to rapidly innovate and bring creative server solutions to the market has enabled Supermicro to provide high-density Twin servers and motherboards with the fastest possible PCI-Express (16) Gen 2 interface to facilitate a maximum bandwidth Tesla server."

“There has been much discussion about the promise of GPU computing,” said Andy Keane, General Manager of GPU Computing for NVIDIA. “This project will demonstrate that Tesla and the CUDA architecture have made the promise a reality. Daresbury is tackling some of the most fundamental scientific questions and we’re confident that our work with the Laboratory and Streamline Computing will play a significant role in answering these challenges.”