Numascale has launched NumaConnect, a scalable shared-memory system built from standard servers. It can be used as a single system or partitioned in smaller systems where each one runs one instance of the OS.
The single memory image cluster provides both shared memory — including threads and OpenMP and MPI programming options. The scalable system takes advantage of low-cost commodity x86 hardware and NumaConnect to offer significant savings compared to conventional shared memory systems.
NumaConnect works with AMD Opteron-based servers and provides up to 256 TBytes of system-wide shared memory using cache coherency logic with a directory-based protocol that scales to 4,096 nodes. The cache coherency logic is implemented in an ASIC together with interconnect fabric circuitry with routing tables for multi-dimensional Torus topologies. This type of fabric is very scalable and the same topology is used in many of the world’s largest supercomputers.
The prototype system that has just finished validation is installed at the University of Oslo. It consists of 72 IBM x3755 servers connected in a 3d torus with NumaConnect. The system runs a standard single image Linux OS on commodity x86 based servers. The cluster is installed under a prototype Partnership for Advanced Computing in Europe (PRACE) study assessing emerging new technologies for European HPC.
The company states that using Numa-awarness applications with high bandwidth requirements it will be possible to utilise the combined bandwidth of all the memory controllers while maintaining low latency access through the shared memory.