Since 1986 - Covering the Fastest Computers in the World and the People Who Run Them

Language Flags
October 29, 2012

How Do You Get To 1TB/s? Quality.

Nicole Hemsoth

Gone are the days when the architects of High Performance Computing (HPC) environments can treat the collection of servers, storage, networking and file systems as a science project for experiments, frequent failures, adjustments and course corrections.

Lowest-cost, commodity-hardware approaches to infrastructure design may have worked at one time, but they are no longer adequate. While, component costs still matter, time-to-productivity and time-to-value are increasingly important and HPC professionals are embracing more turnkey solutions engineered for HPC.  

As HPC applications become more integrated into real-time decision making in government, research and commercial processes, system availability and serviceability have increased in importance, so, too, have the demands for infrastructure reuse, massive scaling and less-disruptive upgrade paths.

IDC stated in their recent document, HPC Storage Systems Target Scalable Modular
Designs to Boost Productivity: June 2012, IDC #233355, “The evolution of HPC storage systems has lagged behind compute server progress, (but a) growing number of HPC storage systems are, or will soon be, pre-integrated, pre-tested, production-oriented solutions.”   Budget, time, resource and environmental constraints require new approaches to HPC storage system design. Xyratex, with the ClusterStor family of integrated HPC data storage solutions, eliminates constraints and places organizations on a path to rapid deployment, increased uptime, infrastructure reuse, linear scaling, environmental efficiency and exascale capabilities.

Quality and Reliability
Due to the massive scale of HPC environments, even a small improvement in component quality translates into higher levels of system availability. HPC storage systems today contain hundreds to thousands of ever-more-densely packed, rapidly-rotating hard disk drives that vibrate, generate heat, consume enormous power and cooling resources, and emit interference that can affect system availability and reliability. When multiplied across the number of system components, Xyratex’s quality improvements dramatically reduce the incidents of out-of-box and production failures.

Xyratex’s approach to storage system design begins with the Xyratex Enterprise Quality Framework (XEQ) and patented four-stage Integrated System Testing Platform (ISTP) that identifies and removes hidden quality problems and significantly reduces in-the-field component failures.  This approach minimizes the time from installation to production and ensures the highest quality storage systems that remain up and available.

Even with improvements in quality, components will, from time to time, fail. The ClusterStor architecture eliminates single points of failure by providing embedded, dual redundant Lustre object store servers, dual redundant management networks, dual redundant Infiniband (IB), 10GbE or 40GbE networks, and one to six ClusterStor 6000 Scalable Storage Units (SSUs).

ClusterStor has been designed from the ground up as a highly available enterprise-class storage solution with resiliency engineered into all primary subsystems. Meta-data and management server operations are resident on a high availability Cluster Management Unit. Each SSU includes active/active embedded server modules, with redundant and independent system interconnects.

Management and Serviceability
Beyond quality and architecture, careful thought to monitoring, reporting, and serviceability are key to minimizing the impact of component failures. ClusterStor Manager provides consolidated health monitoring, error reporting, and diagnostic capabilities for hardware and software components of the entire ClusterStor solution.  ClusterStor Manager removes the management burden of administering tens of petabytes of storage, and their traditional complement of deployed Lustre Object Storage Servers.

To maximize serviceability, each hardware component, whether a power supply, drive unit, or processing unit, is designed for easy access and simple replacement or upgrade, as shown in this video. Component replacements and upgrades can be performed on-line without impacting system availability.

Reuse and Linear Scalability
System reuse and scalability are core design criteria for Xyratex  ClusterStor family of integrated HPC data storage solutions. Customers of the ClusterStor 3000 can scale linearly in capacity and performance by adding additional Scalable Storage Units (SSU).  If requirements exceed the maximum throughput capabilities of the 3000 series, customers can double performance in place, by swapping out ClusterStor 3000 Embedded Server Modules (ESMs) for ClusterStor 6000 ESMs with FDR Infiniband, and preserving drive and rack investments.  The ClusterStor 6000 series, which uses the same software stack and management solution as the 3000 series, scales in 6GB/second increments, providing 42GB/second per rack, enabling installations of up to 1 TB-per-second system throughput, and delivering tens of petabytes of storage.

The Importance of Quality at Scale

Most automobiles driven at modest speed will perform adequately. Generally, if a component fails at low speed, the automobile can be controlled and brought to a safe stop. But when an automobile is driven at racing speeds, small quality issues and component failures can have disastrous consequences. Storage solutions for HPC applications are no different. At the extremes of scale, quality matters more than ever.

Xyratex’ attention to architectural balance and keen eye to detail has resulted in the ClusterStor scale-out data storage solution exactly matching the linear file system processing scalability and quality needs for state of the art HPC scale-out compute architectures.  While alternative scale-out data storage designs have appeared on the market, none match the degree of linear scalability and performance-capacity density delivered by the ClusterStor family of integrated HPC data storage solutions.

Availability, performance, and scale are the three key factors in HPC success. Each factor is directly dependent on quality. Before designing your next cluster gain insight into industry-leading quality processes to make sure your selection includes the attributes and criteria available through the Xyratex Enterprise Quality (XEQ) Framework. To learn more about XEQ click here.