ASTRON & IBM Center for Exascale Technology

RT Communication

Data motion between computers is usually fraught with a large latency caused by multiple copy operations of the data message. In the input-output and communications standards community, over the last couple of years, a standard has been worked out to reduce the number of copy operations to zero, basically by writing the data from the originating computer directly in the application space of the receiving computer.  It is obvious that this can result in significant performance increases. In particular, does this new standard also provide semantics that I.O operations and application processing can be overlapped – a form of parallel processing. Applications which want to make use of both of these advantages need a –partial- re-write, or at least improved Operating System support. The objective is to create a demonstrator computing architecture being able to real-time process high bandwidth, compute intensive workload on an Exascale-class computer. The project aims at (1) overtaking the current limits of I/O subsystems to enable Exascale computers for processing application data at very high bandwidth, and (2) the application of industry standard high performance communication environments for real-time Exascale data processing.

Initially we will identify the systems performance bottlenecks for processing the current workload, and to investigate the applicability of novel, efficient communication models like RDMA (Remote Direct Memory Access) to optimize data I/O-operations and inter-node communication.

In a second phase, the project will investigate the applicability of an highly efficient mapping of industry standard RDMA services onto a specific interconnect network

A third project phase comprises the joint development of a functional prototype system for ASTRON/SKA SKA applications.