In just eleven days during 2007, the Stanford University High-Performance Computing (HPC) Center nearly doubled the performance of its existing compute system. The center leveraged certification methodology from the Intel Cluster Ready Program to fully implement a 1,696-core cluster solution. The solution integrates Clustercorp, Dell and Panasas technologies to give the center the flexibility to meet ever-expanding computational and application requirements and to enable Stanford researchers to achieve faster time-to-results. Steve Jones, the founder and manager of the Stanford HPC Center, writes about the design and deployment of this system.
Mission: Enable CFD on Demand
The goal of our expansion was simple: acquire sufficient compute power to facilitate School of Engineering coursework and research efforts and to support the university’s industrial affiliates program.
The system would need to support more than 200 researchers and effectively enable CFD on demand. Two departments in particular require large-scale, massively parallel computing resources for their work. Researchers in the mechanical engineering and aeronautics and astronautics departments leverage HPC Center resources to analyze the details of flow and acoustics created by helicopters in forward flight. Critical applications include two major in-house-developed simulation codes: Stanford University multiblock (SUmb) and CDP, named for the late Charles David Pierce. Commercial applications include ANSYS, Gaussian, MatLab, and VASP. Stanford’s post-processing programs include EnSight for distributed rendering and Tecplot for visualization.
Objective: Extend In-House Support for Complex Code Verification and Validation
To address Stanford researchers’ increasing needs for code validation, our team opted to bolster local compute resources. The goal was to build a cluster-based solution capable of scaling to thousands of nodes and supporting our large user base. The minimum expectation for the cluster was to run routine jobs in-house and to allow researchers to do more extensive verification and validation of code destined for the national labs. In addition to the large compute capacity and scalability requirements, the system needed to be capable of sustained performance with a file system fast enough to cope with massive I/O load and to deliver the granularity of results researchers require.
Challenge: Overcome Limitations in Space, Time, and Staff
While our plan was to dramatically expand compute resources, the project was bound by constraints of space, time and staff. The Stanford HPC Center has limited square footage in which to operate, so footprint was an issue, not only in terms of the system size and density, but also in how it would impact general HVAC requirements. Extensive project demands also applied scheduling pressures that propelled the deployment team to implement a more aggressive rollout than the three- to six-month timeframe more typically allotted for implementations of this scope. The final challenge was the center’s requirements to grow overall services delivery while maintaining the existing support staff.
We felt that taking advantage of an integrated solution would help us meet our compute and deployment objectives without exceeding our space, time, and staffing limitations. In particular, we were sold on the idea of the Intel Cluster Ready (ICR) program where certification of the Clustercorp/Dell/Panasas solution had been completed upfront. The certification provides cluster components validation to ensure accurate configuration, optimal performance, and a system that is easy to manage and seamlessly expand.
Solution: Use Certified Best-of-Breed Compute, Distribution, and Storage Components
With the assistance of the Dell Advanced Systems Group and following extensive review of available technologies, we began to solidify plans. Specifically, the Dell team worked with us to design and build the final cluster-based solution architecture that incorporates Dell Power Edge Servers using Intel Quad-Core Xeon processors and compilers with Clustercorp Rocks+ cluster distribution software, and Panasas ActiveStor Parallel storage. The design includes:
- Dell PowerEdge 1950 servers. Standards-based product, a 1U form-factor, remote management and diagnostic functionality, scalability, and parts availability were important factors in the selection of Dell servers for the HPC Center compute nodes.
- Cisco 7024 InfiniBand switch. High-speed, full nonblocking internode communications help ensure that compute nodes do not waste cycles waiting for messages.
- Panasas ActiveStor AS3000 parallel storage (with object-based Panasas PanFS parallel file system). Determining factors included Intel Cluster ready certification (Panasas ActiveStor Storage Clusters were actually the first parallel storage solution to be certified as part of the ICR program), MPI I/O optimization, and bandwidth per storage shelf.
- Clustercorp Rocks+ 4.3 (with CentOS 4.5) and Clustercorp Intel Developer, Fluent, Moab, Cisco OFED, and Panasas Rolls. Rocks+ cluster distribution offers comprehensive application integration packages (i.e. Rocks+Rolls) that are essential for large-cluster configuration. Using Rolls to add software stacks with “checkbox” application distribution is a nice relief for system administrators who remember the old days of installation, distribution, configuration, scripting, and testing for every application added to the system. Rolls for Fluent, EnSight, and other higher-level applications greatly add to the simplicity of building an end-to-end solution.
- Dell Deployment Services. The Dell Advanced Systems Group and the company’s ecosystem of partners, including APC, Cisco, Clustercorp, Intel, and Panasas, came together to architect and implement the solution seamlessly and in record time. Expertise and comprehensive planning were critical to the rapid deployment.
- APC Hot Aisle Containment System. This chilled-water, row-cooling system enables greater rack density (and therefore smaller footprint) than would be possible using traditional room cooling.
In terms of the selection of the storage component, let me add further explanation of our selection of the Panasas parallel solution. Unlike NFS appliances and other clustered-storage products we considered, the Panasas storage system enhanced application performance to deliver the parallel I/O efficiency and stability our researchers require. Our experience with NFS has been that as we start to increase the number of processors or processes writing in parallel to the file system, we overwhelm the appliance and cannot successfully run a large simulation. That’s why we looked for a storage solution specifically optimized for parallel processing environments.
We did run direct comparisons of Panasas PanFS parallel file system against an array of other file systems, including those commonly targeted to HPC clusters. The PanFS parallel file system consistently outperformed them–in fact, when we ran our simulation and modeling applications, Panasas parallel storage allowed it to run significantly faster. We’ve used Panasas parallel storage in-house for more than three years now because we find it to be the highest-performing, most manageable, and most reliable of the leading storage solutions designed for high-performance computing.
Result: Achieve 2-14X Performance Improvement After an 11-Day Deployment
The entire deployment, including implementation of an entirely new power and cooling infrastructure, took a total of eleven days. The Dell Enterprise Deployment team played an integral role in this feat, coordinating the efforts of all participating vendors. The power, cooling, and system build-out were completed in parallel. We used the Rocks+ Linux cluster distribution to configure master and compute nodes, and by day eleven our researchers were able to submit jobs that were flawlessly executed and producing scientific code and operations with unprecedented fidelity. The new cluster easily handles ten times the workload of the original 48-node configuration. Testing results show performance of 15.8 teraflops performance compared to 1.1 teraflops delivered by the smaller cluster.
The Stanford HPC Center currently supports fifteen different types of HPC systems, including the Intel Cluster Ready system based on the Clustercorp/Dell/Panasas solution. The original 48-node system is now used by students and researchers working on smaller-scale problems. The ease-of-use of commodity-based hardware, efficient cluster distribution, and appliance-like ease of storage management allow the center to maintain a small support staff.
Research Impact: Achieve Faster and More Accurate Time-To-Results
The new cluster is paying off well for our center. It has delivered faster time-to-results by:
- Allowing researchers to run more simulations and other heavy-compute jobs in-house.
- Delivering finer detail or higher fidelity so that scientists more quickly and easily recognize salient features or other important phenomena.
- Enabling more thorough verification and validation of complex codes destined to be submitted to the 10,000-65,000-processor systems at the National Labs.
With this new, more powerful, stable and manageable cluster in place, our researchers are better able to focus on their science and deliver consistent, meaningful results to project sponsors. Deployment of the new cluster has also brought added recognition to the Stanford High Performance Computing Center, helping us achieve a ranking of 130 on the Top 500 list in November 2007.
Meanwhile, our compute infrastructure continues to grow — the HPC Center is currently doubling the size of the existing cluster and designing an additional system based on a similar architecture. The integrated, standards-based solution and the Intel Cluster Ready certification combine to give the Stanford High Performance Computing Center the flexibility to add or change cluster elements to meet specific computational and application requirements. Stanford’s industrial partners are also able take advantage of the system to enable more traditional operations and expand their own research and computing services.
About the Author
Steve Jones currently runs the High Performance Computing Center at Stanford University, supporting sponsored research for The Department of Energy Advanced Simulation and Computing Program (ASC), and the next-generation Predictive Science Academic Alliance Program (PSAAP). The HPC Center also supports the computational needs of sponsored research for National Aeronautics and Space Administration (NASA), Air Force Office of Scientific Research (AFOSR) and Defense Advanced Research Projects Agency (DARPA). Jones is the chair of the annual Stanford High Performance Computing Conference, has designed and currently administers numerous Top 500 Supercomputers, and speaks regularly about the management of High Performance Computing Clusters. More information can be found at http://hpcc.stanford.edu and http://psaap.stanford.edu.