|Article title||SPEEDUP PREDICTIONS FROM HYBRID RECONFIGURABLE SYSTEMS|
|Section||SECTION I. PRINCIPLES OF CONSTRUCTION, ARCHITECTURE AND HARDWARE BASE SUPERCOMPUTERS|
|Month, Year||12, 2014 @en|
|Abstract||This paper presents speedup predictions for a reconfigurable hybrid computing system containing a fixed number of processor cores and accelerators and switching components to vary connections between the processor cores and accelerators. Maximum acceleration criteria for given values of initial process parameters are specified. As distinct from FPGA-based reconfigurable systems, the systems considered in this paper contain only two components: a MIMD-component, being a set of universal processor cores, and a SIMD-component, being a set of arithmetic accelerators. Restricting the types of components to these two computing structures makes it possible to considerably simplify the algorithm of decomposition and reconfiguring. In particular, it resolves into (hardware- or software-) switched connection of processor cores and accelerators implemented based on the analysis of the values of initial process parameters. The relationships obtained show that if certain conditions are fulfilled, reconfiguring makes it possible to speed up the process of computations by hybrid systems significantly, but no more than by a factor of two, in comparison with non-reconfigurable systems. Vendors of computer components have established production of systems structurally identical to those studied in this paper. Application of the solutions proposed in the paper renders computing systems more efficient.|
|Keywords||Hybrid computing system; MIMD-component; SIMD-component reconfigurable system.|
|References||1. Stepanenko S.A. Otsenki uskoreniya vychisleniy gibridnymi sistemami [Evaluation of the acceleration of hybrid computing systems], Plenarnye doklady Pyatoy Mezhdunarodnoy
konferentsii «Parallel'nye vychisleniya i zadachi upravleniya» PACO 2010. Moskva, 26–28 oktyabrya 2010 g [Plenary papers of the Fifth International conference "Parallel computations and control problems" PACO 2010. Moscow, 26-28 October 2010]. Moscow: Institut problem
upravleniya im. V.A.Trapeznikova RAN, 2010, pp. 61-71.
2. Kalyaev I.A., Levin I.I. Rekonfiguriruemye mul'tikonveyernye vychislitel'nye sis-temy dlya resheniya potokovykh zadach obrabotki informatsii i upravleniya [Reconfigurable multiconference computing system for solving ow problems of information processing and
management], Plenarnye doklady Pyatoy Mezhdunarodnoy konferentsii «Parallel'nye vychisleniya i zadachi upravleniya» PACO 2010. Moskva, 26–28 oktyabrya 2010 g [Plenary papers of the Fifth International conference "Parallel computations and control problems" PACO 2010. Moscow, 26-28 October 2010]. Moscow: Institut problem upravleniya im.V.A. Trapeznikova RAN, 2010, pp. 23-38.
3. Bondalapati K., Prasanna V.K. Reconfigurable Computing Systems, Proc. IEEE, 2002, Vol. 90, No. 7, pp. 1201-1217.
4. Stepanenko S.A. Sposob opredeleniya struktury gibridnoy vychislitel'noy sistemy [The way to define the structure of the hybrid computing system]. Patent RF № 2436151 na izobretenie.
5. Oak Ridge changes Jaguar's spots from CPUs to GPUs. Available at:
6. Tsil'ker B.Ya., S.A.Orlov. Organizatsiya EVM i system [Organization of computers and systems]. St. Petersburg, 2004, 668 p.
7. Available at: http://content.dell.com/us/en/gen/ spredir.ashx/solutions/poweredge-c410x-thinkers.