High Performance Computing (HPC)

High performance computing (HPC) uses parallel processing over the fastest compute, interconnect and storage technologies available, in order to simulate and solve scientific problems in the shortest time possible. Such highly intensive computation tasks are common in many scientific fields, including automotive & aerospace design, financial modeling, reservoir simulation, climate modeling and national defense.

Over the past decade, HPC clusters based on standard, “off-the-shelf” components have become the dominant solution for building cost-effective supercomputers with extreme performance and scalability. The HPC clusters that are deployed today run tens of thousands of parallel processes, using high bandwidth, low latency interconnects for efficient communication.

InfiniBand delivers the highest bandwidth and lowest latency of any standard interconnect technology, along with the lowest CPU utilization, making it the ideal interconnect for high performance, parallel applications.  As the recognized leader in InfiniBand switching fabrics, Voltaire InfiniBand solutions provide the perfect interconnect for HPC.

Fast, Scalable Interconnect

Voltaire InfiniBand scale-out data center fabrics feature high-bandwidth, low latency smart switches with the ability to run mesh topologies and multiple paths. Voltaire InfiniBand solutions allow horizontal scalability without performance degradation, while maintaining a linear cost model – making them ideal for large HPC clusters.

The smart design of Voltaire’s 40 Gb/s (QDR) InfiniBand switches provides unprecedented levels of performance. The switches can be combined to form highly scalable, cost effective, low latency fabrics. As a result, I/O bottlenecks are removed, allowing applications to operate at maximum efficiency. Additionally, using Voltaire’s unique HyperScale architecture, it’s possible to build clusters that can scale out to thousands of nodes with a single tier of switches and under 400 nanoseconds of latency. 

HyperScale Mesh 1296-node example compared to the traditional approach

Application Acceleration Software

Group communications challenges are common in high performance computing applications such as reservoir modeling, fluid dynamics, and crash analysis. Voltaire enables HPC users to accelerate MPI collective operations by offloading significant parts of group communication onto the switching fabric and orchestrating an efficient, topology-based collective flow. This fabric-wide solution leverages both Voltaire’s fabric management software and its smart InfiniBand switches, which incorporate a CPU on every module, eliminating congestion and reducing latency throughout the fabric. As a result, bottlenecks are removed at both the server and interconnect levels, resulting in 10X to 100X collective performance acceleration and a 10% to 40% improvement in overall application runtime.

Fluent benchmark results showing application runtime acceleration using Voltaire FCA fabric collective offloading technology

Additionally, Voltaire’s fabric management solution can be used to optimize routing within the fabric and thus improve workload performance. By taking into account end-to-end job traffic patterns, Voltaire’s unique routing optimization algorithm reduces competition between resources and increases available bandwidth to maximize utilization of the fabric.

10GbE for HPC

While InfiniBand continues to be the leading interconnect for HPC environments, more conservative users will always prefer to stick to “good-old Ethernet,” especially for applications that are less communication-intensive. With 10GbE latencies and falling prices, along with the emergence of 10GbE LOM (LAN on Motherboard) technology from server vendors, Ethernet is now a viable interconnect for many HPC environments. Voltaire’s 10GbE solutions enable users to benefit from a far more scalable, lower latency, and virtualized fabric with lower overall fabric costs and power consumption, greater efficiencies, and simplified management than traditional 10GbE fabrics. For a customer building a 1000 node data center, this approach delivers 10X lower latency and 4X faster core performance for half the price, using 3X less power than alternative 10GbE solutions.

Field Proven, Fully Integrated Solutions

Voltaire solutions are fully integrated with the HPC offerings of all leading server OEMs. With years of experience working with large HPC customers, Voltaire has a deep understanding of the challenges HPC users face from both an interconnect and an application perspective, and its HPC solutions are backed up by Voltaire’s industry-leading support organization. Voltaire’s solutions are also integrated into leading cluster management, scheduling, and OEM solutions through an open set of APIs and tools.

Voltaire solutions have been at the top of both the TOP500 (top500.org) list of fastest supercomputers in the world for years, indicating recognized leadership in both performance and efficiency. On the TOP500 list published in June 2010, 55% of the InfiniBand deployments on the list used Voltaire switches – double the share of any other InfiniBand systems vendor.