Technology Prevalence in the HPC Market September 2007

EXECUTIVE SUMMARY

This Tabor Research report describes technology options available to users in the High Productivity
computing market. For this report, Tabor Research reviewed specifications for 153 server products which
are actively sold into the HPC market by 14 suppliers. We collected technical data at both the system level and the node level. We reviewed 177 nodes for the study.

We segmented systems by architecture type, including clusters, blades, MPPs, and SMPs, and we compared
the availability and prevalence of component technologies by type. The technologies we reviewed included
system interconnects, operating system options, and processors.

We segmented nodes by form factor type, including rack-mounted, blade, tower, and standalone. The
availability and prevalence of component technologies were compared by type. Technologies reviewed on
a per-node basis included sockets, cores, memory, disk, I/O ports, and processors. Node-level analysis
included standalone systems as “single node systems,” and where appropriate we made share calculations
both with and without the standalones.

Key findings from the study include:

  1. Rack-mounted nodes account for over half of all the node options available in the market today.
  2. Nodes average about three sockets per node when standalone systems are excluded. The
    standalone products average approximately 55 sockets.
  3. Maximum memory per node (excluding standalone) averages approximately 51 GB.
  4. PCI-X ports were the most common, averaging 1.3 per node (excluding standalone), followed by
    PCI Express, which averaged 1.2 ports per node.
  5. The vast majority of nodes – about 66% – used x86-64 architecture processors.
  6. AMD Opteron processors and Intel Xeon processors accounted for virtually the same number of
    available nodes.
  7. Ethernet interconnects were available on 58% of distributed memory systems, followed by
    Infiniband with about 22%.
  8. The Linux operating system is available in some form on virtually all systems. Windows OSs are
    available on over half of HPC systems.

We looked at the differences between blades and rack-mounted nodes in some detail. There are about three
times as many rack-mounted products available in the market as blade products. Their respective node
strategies differ in two major areas:

  1. Maximum capabilities – Rack-mounted alternatives provide advantages in such areas as sockets
    per node, disk drives per node, maximum memory, and number of I/O ports. In large part these
    differences result from the original design objectives for blades, which focus on minimizing space
    and power requirements. That said, there is no inherent technical barrier preventing vendors from
    creating large form factor blades that would have similar capabilities to rack-mounted nodes.
  2. Added value versus standardization – Blades provide added value over rack-mounted nodes
    specifically due to their differentiated architecture. These systems provide a number of features
    designed to simplify system integration, increase system reliability, and decrease environmental
    costs. However, these features are added at the cost of “de-standardizing” the product.

This analysis of cluster technology has led us to look at alternative approaches for segmenting or
classifying computer system technology. We believe current computer architectures are best classified
along two dimensions: memory architecture (i.e., logically shared memory models versus distributed
memory models) and degree of product standardization.

Clear