Soft Coherence : Preliminary Experiments with Error-Tolerant Memory Consistency in Numerical Applications

As we scale into the multi-core era, we face severe challenge s in the scalability and performance of on-chip cache-coherent shared memory mechanisms. We exp lor application error-tolerance as an extra degree of freedom to meet these challenges. Iterative numerical algorithms, in particular, can cope with the occasional stale value with little or no effect on ac curacy or convergence time. We explore analysis methods to distinguish between critical and non-c ritical data in such algorithms. We exploit this distinction to designsoft coherenceprotocols that provide strong guarantees for critical data and weak guarantees for non-critical data. Our preliminary results se a conjugate gradient solver as an example, with experiments on five sparse matrices showing 6.9%-12.6% performance improvement, with little loss in precision.