Cray XT4 and Seastar 3-D Torus Interconnect

The physical sciences are increasingly turning toward computational techniques as an alternative to the traditional “wet lab” or destructive testing environments for experimentation. In particular, computational sciences can be used to scale far beyond that of traditional experimental methodologies; opening the door to large-scale climatology and molecular dynamics, for example, which encompass enough detail to accurately model the dominant terms that characterize the physical phenomena being studied [2]. These large-scale applications require careful orchestration among cooperating processors to ply these computational techniques effectively. The genesis of the Cray XT4 system was the collaborative design and deployment of the Sandia “Red Storm” computer which provided the computational power necessary to assure safeguards under the nuclear Stockpile Stewardship Program which seeks to maintain and verify a nuclear weapons arsenal without the use of testing. It was later renamed the Cray XT3 and sold commercially in configurations varying from hundreds of processors, to 10s of thousands of processors. An improved processor, faster processornetwork interface, along with further optimizations to the software stack and migrating to a lightweight Linux kernel prompted the introduction of the Cray XT4; however, the underlying system architecture and interconnection network remained unchanged.