By Martin Gerber
Read Online or Download Parallelising Molecular Dynamics for Message Passing Systems (Berichte Aus Der Informatik) PDF
Similar mathematics books
- Algebraic Function Fields and Codes (Graduate Texts in Mathematics, Volume 254)
- A=B (symbolic summation algorithms)
- Elements de Mathematique. Algebre commutative. Chapitres 5a7
- Differential Equations in Banach Spaces: Proceedings of a Conference held in Bologna, July 2–5, 1985
- Jeux de l'esprit et divertissements mathematiques
Extra info for Parallelising Molecular Dynamics for Message Passing Systems (Berichte Aus Der Informatik)
The list scheduling algorithm is only capable to schedule non-cyclic dependence graphs into non-iterative schedules. Iterative schedules are characterised by different values for latency and period: the latency is the time which is needed to process the dependence graph one time. If the dependence graph is invoked repetitively, a result is delivered after each period. With a good pipelined schedule, period can be significantly smaller than the latency thereby improving the throughput. The basic list scheduler had been extended such that iterative algorithms are planned with minimised period.
These numbers denote the speed-up of the tested machine compared with a reference machine. The SPECx95 reference machine is a Sun SparcStation10 (40MHz). 30 Chapter 2: Methodology Although it is possible to predict the execution time of a custom application through the use of SPEC results, it is preferable to measure the delivered performance directly at runtime. If an application (or function) specific delivered performance of the target machine is determined, Perfj(vi) represents the performance for a given function vj executed on the computational node vi.
2 Specification and Modelling 29 To keep our model as simple as possible, a fixed message length per communication node wi was assumed so that the latency and the bandwidth are independent of the message length. 1, with the following restriction: broadcast and collective communications (reduce) are always modelled with successive point-to-point links, with messages of a constant length wi. 2). Furthermore it was assumed that the communication of the previous function may always overlap the computation of the current function.