Exploring Scalability Techniques of OmpSs
Iulian Brumar · RECERCAT (Consorci de Serveis Universitaris de Catalunya) · 2014
Back to the 70's the Moore's law was launched with unstoppable force until 15 years ago. The rule says that the number of transistors per area unit will double every two years. One benefit of this rule was that the pro- grammers only had to wait two years and their programs went the double as fast. However after the year 2000, the programmers have to do significant ef- fort to make their programs run faster, and eficient programming techniques became more and more important. In a multi-core era, parallel programming allows further performance im- provements, but with an important programmability cost. We envision that the only approach to parallel programming that can exceed the program- ability, power, memory and reliability walls in Computer Architecture is a run-time approach, which will allow riding once more on Moore's law. The runtime approach revives 80's and 90's computer architecture for- gotten concepts that can be applied at the runtime layer in a completelly transparent way to the programmer. The goal of this work is taking the computer architecture value prediction concept inside a runtime environment like OmpSs. Learning from the execution history, allows the runtime-placed value pre- dictor to obtain a 2x average speedup for three kernels with fine-grained tasks: blackscholes, checksparseLU and jacobi.