A Loosely Coordinated Model for Heap-Based Priority Queues in Multicore Environments

An adaptive algorithm for high‐dimensional integrals on heterogeneous CPU‐GPU systems

“…Such results confirm our expectation of acceptable overhead of the redistribution procedure on the efficiency of the algorithm, because there are not global synchronization among the threads. Only families F (3) , F (5) , and F (6) show values for the efficiency E N < 0.7 with N = 16 threads, because the presence of features (the peak in the corner or the discontinuity) needing a more frequent redistribution of sub-domains. A last set of experiment is aimed to measure the effectiveness of the GPU as floating point accelerator in the hybrid Algorithm 2 as described in the previous section.…”

“…More precisely, the families F (2) and F (3) show a performance gain of about 5× because their analytic expressions are based only on floating point operations, without the use of trigonometric or exponential functions as for the families F (1) , F (4) , and F (5) . The worst case is represented by the family F (6) with a performance gain of about 1.8× because the thread divergence due to the presence of the selection structure in its analytic expression, that greatly limits the GPU performance when threads follow different paths in the control flow.…”

“…In a previous work, we have shown that, in the parallelization process of Algorithm 1 in a shared memory environment, two strategies are available: a first approach is based on a centralized data structure for storing the sub‐domains, where several global synchronizations are needed among all threads in order to process only sub‐domains with large error, while in the second approach the sub‐domains are distributed in separate data structures managed by different threads, without synchronizations but with a significant risk that some of them process items with small error. Let us recall shortly the strategy on the basis of an effective trade off between the previous two requirements.…”

“…For example, our previous work introduced an algorithm based on two adaptive strategies on MPP systems with multi‐core CPU, combining together two different programming models for distributed memory and shared memory environments.…”

“…The work presented in the paper is the continuation of our investigations on the development of adaptive algorithm on modern hardware platforms (see, eg, our previous works), where we focused our attention mainly on computing systems based on multi‐core CPUs only. Some preliminary results on a hybrid system are reported in another previous work .…”