Program Scalability Analysis for HPC Cloud: Applying Amdahl's Law to NAS Benchmarks

Scalability Dilemma and Statistic Multiplexed Computing — A Theory and Experiment

“…In Figure 7, the optimal granularity sizes are: 11, 22, and 33. The relationship between these sizes are given by the volatility power indices of the 45 cores at their peaks: 303, 409, and 818 [20]. At these optimal points, the synchronization overhead is near zero (note the Brachistochrone analogy).…”

“…At these optimal points, the synchronization overhead is near zero (note the Brachistochrone analogy). Further quantitative scalability analysis becomes possible using the optimized performance and simpler time complexity models [20]. The MPI (Open MPI 1.10.0) on Centos 7 delivered consistently worse performance than the worst tuned Synergy performance (5.1GFLOPS).…”

“…The optimal processing granularity (G) of an application defines the Termination Time Equilibrium [19], [20] for that application. Finding the optimal G will allow the SMC applications to overcome the poor performance stigma of earlier dataflow and Tuple Space machines.…”

Parallel Timing Model Applied to Hadoop Applications on a Private Cloud

“…On the surface, Amdahl's Law offers a pessimistic view of parallel processing, showing that speedup drops off drastically if even a tiny portion of the program is not parallelizable. [5] For this reason, John L. Gustafson revised Amdahl's law and created a scaled formula, [5] where…”

“…Though the two laws are mathematically equivalent, [5] there is a difference in their applicability. Because Gustafson's Law includes t p (P ) rather than just t p (1) as Amdahl's Law does, it is possible with Gustafson's Law to quantify parallel and communication times in the same percentage.…”