A typical metric used to evaluate network performance is its latency for point-to-point communications. But more important, and sometimes overlooked, is the latency for collective communications, such as barrier synchronization used to synchronize a set of processes, and all-reduce, used to perform distributed reductions. For many High-Performance-Computing applications, the performance of such collective operations play a critical role in determining overall application scalability and performance. As such, a system-oriented approach to network design is essential for achieving the network performance needed to reach extreme system scales.
The CORE-Direct technology introduced by Mellanox was a first step at taking a holistic system view, by implementing the execution of collective communications in the network. The Scalable Hierarchical Aggregation and Reduction Protocol (SHARP)™ technology being introduced is an extension of this technology, which moves support for collective communication from the network edges, e.g. the hosts, to the core of the network – the switch fabric. Processing of collective communication moves to dedicated silicon within the InfiniBand Switch-IB 2 switch, thus providing the means for accelerating the performance of these collective operations by an order of magnitude
This large increase in performance greatly improves the performance of applications and reducing the time to perform operations such as a global reduction used to compute the magnitude of an error vector used by iterative solvers. Such computations consume many wall-clock cycles for many scientific and engineering codes, such as OpenFOAM and others. As such, this technology will increase the abilities of such applications to address urgent business needs.
Next generation HPC infrastructures will depend on smart interconnect and offloading architectures to deliver the needed performance, efficiency and scalability. Switch-IB 2 delivers the best network infrastructure, bringing us one step closer to Exascale computing.