HP XC6000 Cluster at SSCK

At the SSCK the installation of an HP XC6000 Cluster as high performance computer of the state has started with the beginning of 2004. It will be set up in several stages. The system, operated under Linux, consists of powerful single nodes which contain each two, four or 16 Intel Itanium 2 processors being linked by a fast interconnect (Quadrics QSNet II).

In the final stage of expansion in the beginning of 2006 the redundantly designed, highly available system with 1,200 processors will achieve a computing power of 11 TFlop/s and provide more than 7 TB of main memory.

After the installation of the test cluster with 16 nodes in April 2004 the system expansion of the first phase (end of 2004 / beginning of 2005) will dispose of a peak perfomance of 2.2 TFlop/s as well as a memory extension of 2.2 TB.

Structure of the HP XC cluster in phase 1:

• 116 two-way nodes with 12 GB main memory each
• 6 sixteen-way nodes with 128 GB main memory each
• Single Rail Quadrics QSNet II Interconnect
• 10 TB global disk space

In phase 2 (beginning of 2006) this system will be extended by:

• 218 four-way nodes with 24 GB main memory each
• Dual Rail Quadrics QSNet II Interconnect
• 30 TB global disk space

Fast Communications Network and High Scalability

The architecture of the HP XC Cluster is characterized by a clear structure and specialization of the single nodes. Two-way nodes respectively four-way nodes are used for applications being parallelized by MPI. The communications network with a latency of about 3 µs and a bandwidth of about 800 MB/s on MPI level allows a high scalability so that also communication intensive applications with high processor numbers can be performed efficiently.

Applications being parallelized according to the principle of shared memory access can be carried out on the sixteen-way nodes and benefit from the common main memory with 128 GB and the local disk capacity of more than 1 TB per node. The utilization of the sixteen-way nodes is also planned for interactive applications for pre- and postprocessing as well as for data filtering.

Parallel Cluster File System Lustre

With Lustre the HP XC6000 Cluster possesses a global parallel file system being designed for very large clusters and high I/O bandwidths. By the utilization of several object storage servers (OSS) und meta data servers (MDS) a parallelism of data access and also a redundancy in case of a failure of single servers is achieved.

In the first expansion stage 10 TB, in the second stage 40 TB will be available for global file systems. Moreover, every node of the XC Cluster is provided with local disks for temporary files.

High Efficiency by Cache Utilization

The nodes of the HP XC Cluster are based on Intel Itanium 2 processors. These processors are in particular characterized by a high performance in the field of floating point arithmetic as well as a very large date cache, which is located on the processor chip and therefore can be addressed with a very short latency and extremely high bandwidth. Thus the system is especially suited for application programs being optimized for cache utilization.

Simple Porting of Application Programs

The programming environment and application interfaces of the HP XC6000 system are based on open standards and therefore permit a simple porting of application programs.