UCAM 042/14 For the provision of design and optimized software for Square Kilometre Array Computatio

Descriptions

The Square Kilometre Array (SKA, http://www.skatelescope.org/) is a major international project to build a next generation radio astronomy observatory. The anticipated construction cost of the full SKA is about 2 billion EUR. When completed, the SKA will be, by a large margin, the largest and most sensitive low-frequency radio observatory ever built.The design of the main elements of the SKA is being carried out by International Consortia, with the SKA Organisation coordinating the design process. The SKA Organisation is also the overall design authority for the SKA. The SKA Organisation is a UK company limited by guarantee and is based in the UK at Jodrell Bank in Cheshire.An integral part of the observatory is processing the data from the streams produced by digital signal processing into calibrated, usable science products. This processing is the responsibility of the “Science Data Processor” (SDP) element of the SKA which is led by the University of Cambridge. The “Science Data Processing Consortium” includes members on five continents and a mixture of Universities and government organisation and institutes.The deployment of the telescopes in SKA will be phased, with the first phase (SKA1) due to commence procurement in 2017. Until 2017 the SKA project is in a pre-production phase during which elements are undergoing design and technology is brought to production readiness level. The observatory will consist of telescopes which will be sited in Southern Africa and Western Australia. There will be a computing facility for Science Data Processing on each of the continents.The purpose of the work to be awarded is to assist the University of Cambridge in its role as a part of the SKA SDP consortium in understanding the potential role of massive SIMT and SIMD-type accelerators such as GPUs and Xeon Phi (accelerators) in the SKA SDP compute hardware platform. In order to achieve this, the University of Cambridge seeks services which allow the consortium to:— Understand the current, and likely future, absolute performance and performance per Watt of key algorithmic steps for radio-interferometric data reduction on high-throughput general-purpose accelerators likely to be on the market from 2016 onward;— Obtain implementations optimised for accelerators of mathematical kernels while:Integrating and co-designing dataflow interfaces which the SDP is developing and;Implementing API's for the kernels that are portable to multiple architectures.— Examine in conjunction with other individuals and groups in the SDP consortium how data flow to and from the kernels can be optimized and scaled to handle relevant data sets;— Understand and characterize the factors limiting the absolute performance and performance per Watt of such kernels.The University of Cambridge contact person will provide guidelines for the data driven frameworks, pseudo-code and in some cases C/C++/CUDA/ OpenCL/Fortran code of the key computational steps. A preliminary SKA SDP System Description [RD 8] is made available with this invitation together with an analysis of computational requirements [RD 9]. [RD 8] and [RD 9] are preliminary documents intended to give the context of the work being procured here.