December 5, 2011

Selecting the Most Effective InfiniBand Topology for Technical Computing

Nicole Hemsoth

SGI® ICE 8400

Selecting the Most Effective InfiniBand Topology

Across a wide range of disciplines, InfiniBand technology now enables clusters that range from a few systems to the largest technical computing clusters in the world. In only a few years, clustering with InfiniBand has come to easily dominate the top 100 of the Top500 list of supercomputing sites ( As new grand-challenge problems and other computational challenges emerge, larger and larger clusters will be required. Even with now routine advances in processor speed and memory capacity, scaling with cluster size will likely remain the simplest way to grow computational capacity for the world’s most tenacious computational problems. While InfiniBand can be deployed in multiple topologies, choosing the optimum InfiniBand topology can be difficult, with trade-offs in terms of scalability, performance, and cost. SGI has considerable experience in the design and deployment of some of the largest InfiniBand clusters in existence.

While some vendor’s limitations drive them to push one topology choice above others, SGI understands that the best topology is one that matches the needs of the application. Based on high-performance AMD Opteron™ 6200 Series processors, the SGI® ICE 8400 system is designed for flexible and optimized InfiniBand topology configuration.

InfiniBand Topology Considerations and Trade-offs

SGI ICE supports multiple InfiniBand topology choices, including All-to-All, Fat Tree (CLOS), as well as Hypercube and Enhanced Hypercube topologies. Choosing the right topology involves understanding the needs of the application as well as comparing key metrics and cost implications.

SGI ICE Topology Choices

InfiniBand fabrics present different advantages and limitations. The SGI ICE system is designed to flexibly support multiple InfiniBand topologies, including:

  • All-to-All. All-to-All topologies are ideal for applications that are highly sensitive to Message Passing Interface (MPI) latency since they provide minimal latency in terms of hop-count. Though All-to-All topologies can provide non-blocking fabrics, and high bisection bandwidth, they are restricted to relatively small cluster deployments due to limited switch port counts.
  • Fat Tree. Fat Tree or CLOS topologies are well suited for smaller node-count MPI jobs. Fat Tree topologies can provide non-blocking fabrics and consistent hop counts resulting in predictable latency for MPI jobs. At the same time, Fat Tree topologies do not scale linearly with cluster size. Cabling and switching become increasingly difficult and expensive as cluster size grows, with very large core switches required for larger clusters.
  • Standard Hypercube. Standard Hypercube topologies are ideal for large node-count MPI jobs, provide rich bandwidth capabilities, and scale easily from small to extremely large clusters. Hypercubes add orthogonal dimensions of interconnect as they grow, and are easily optimized for both local and global communication within the cluster. Standard Hypercube topology provides the lightest weight fabric at the lowest cost with a single cable typically used for each dimensional link.
  • SGI Enhanced Hypercube. Adding to the benefits of Standard Hypercube topologies, SGI Enhanced Hypercube topologies make use of additional available switch ports by adding redundant links at the lower dimensions of the hypercube to improve the overall bandwidth of the interconnect

SGI ICE 8400: Designed for InfiniBand

The SGI ICE platform is fundamentally architected to provide cost-effective high-performance InfiniBand infrastructure. The SGI ICE 8400 platform in particular is capable of achieving industry-leading scalability without sacrificing application performance efficiency. The platform offers a variety of interconnect options that let organizations scale their applications across hundreds or thousands of processor cores.

The SGI ICE 8400 system can accommodate up to 16 compute blades within each Individual Rack Unit (IRU). The

IRU is a 10 rack unit (10U) chassis that provides power, cooling, system control, and network fabric for up to 16 blades via a backplane. Up to four IRUs are supported in each custom-designed 42U rack, with a choice of either air cooling or water cooling for all configurations. Each rack supports:

  • A maximum of four IRUs
  • Up to 2048 of AMD Opteron ™ 6200 series
  • A maximum of 12.2TB of memory (64 x 192GB)


Effective InfiniBand topology requires system architecture designed with scalability in mind. The SGI ICE system

was purposely designed for InfiniBand networking, and together with the high core density of  AMD Opteron 6200 Series processors, the platform is capable of achieving industry-leading density and scalability for a broad range of technical computing applications. Being the world’s only 16-core x86 processor, the AMD Opteron 6200 Series processor delivers unprecedented scalability for large HPC deployments. With a choice of supported InfiniBand topologies, the SGI ICE system is ideal for deploying InfiniBand clusters ranging from a single 16-node IRU to hundreds of racks and many thousands of nodes.

Selecting an appropriate InfiniBand topology requires careful consideration of applications, algorithms, and data sets, along with likely needs for scalability into the future. In the absence of benchmark data, having some basic knowledge of the application characteristics may be enough to guide topology choices. Extensive testing done by SGI has shown that applications are generally less sensitive to topology than kernel benchmarks, but that differences in performance become more pronounced as clusters grow in size. When global interconnect bandwidth is important, Enhanced Hypercube dual-rail is the raw performance leader. For smaller single-rail topologies, Fat Tree is often the most economical choice. As clusters grow, hypercube topologies gain scalability, performance and cost advantages, avoiding the external switching and cabling that is required for Fat Tree and All-to-All topologies. Having deployed some of the world’s largest open systems InfiniBand networks and clusters, SGI has the experience and expertise to help organizations choose the right equipment and networking topology to meet their most challenging computational problems.

For more information go to:

Share This