Cray XT5h

The XT5h uses a combination of different processor architectures and arrangements to tackle different stages during a problem solution. This hybrid architecture helps the supercomputer take advantage of very fast, very parallel, and very specialized programming techniques in separate parts of the machine.

Key Facts

• The XT5h combined meshes of AMD Opteron processors with tightly integrated groups of vector processors and specialized reprogrammable processors
• Cray code-named the vector multiprocessor chips in the XT5h “Black Widow”
• The XT5h was the first hybrid supercomputer to break petaflop speed
• At the time of its release in 2007, the Cray XT5h was considered the most powerful hybrid supercomputer in the world
• Applications included superconductivity and thermodynamics research at the Oakridge National Laboratory, and cryptographic work at the National Security Agency

The XT5h was a special version of the Cray XT5 and was developed from the XT range of Cray supercomputers. The key feature that made the XT5h different from the XT5 was the addition of vector processor blades and reprogrammable special-purpose processors. The XT5h was the first hybrid supercomputer to integrate multiple processor architectures into a single system. In a Cray XT5h system, combinations of Cray XT4, XT5, XR1, and X2 blades could be integrated. The Cray XT5h also supported global memory programming and the ability to assign software applications to specific parts of the hardware in the system.

Three key parts of the system made it appealing to customers needing fast, high-volume, and customizable computing.

Cray XT4 and XT5 Blades

Blades from earlier Cray XT4 systems could be combined with XT5 blades to provide an upgrade path. The XT5 blades contained two AMD Opteron dual or quad-core processors with dedicated memory and an interconnection Cray SeaStar2 chip.

Cray XT5h Software Features

Special software features ensured that the hybrid hardware worked efficiently, scaled easily, and operated smoothly.

Application-Level Placement Scheduler (ALPS)

The ALPS was a software feature that allowed moving applications to different parts of the machine from within the application. This allowed applications to take advantage of the hybrid environment, where they could be assigned to run on parts of the system best suited to the task.

Lustre File System

The Lustre file system was a parallel file system that scaled to thousands of clients and petabytes of data. Lustre was designed for high performance and scalability. By providing parallel access to storage, many different parts of the processing system could access data simultaneously and then bring parts back together once complete.

Hardware Supervisory System (HSS)

Like an electronic janitor, the HSS made sure everything worked properly by monitoring and alerting when any hardware or software failure was detected. It also enabled reconfiguration and the phased start-up of cabinets and other equipment.

Advanced Cray XT5h Programming Language Features

Advanced parallel programming languages, Co-Array Fortran and Unified Parallel C, allowed programmers to specify data placement locations in a machine. This enabled improved application performance, efficiency, and scalability of applications involving large or complex data sets.