China
Africa
Explore chapters and articles related to this topic
Homotopy Algorithms for Engineering Analysis
Published in Hojjat Adeli, Supercomputing in Engineering Analysis, 2020
Layne T. Watson, Manohar P. Kamat
As stated previously, the percentage of serial execution time that is spent in the evaluation of the polynomial system and its Jacobian matrix ranges from 50% to 80%. The percentage depends on the complexity of the polynomial system. As the complexity increases the fraction that can be parallelized increases. This also increases the granule of parallelization and thus the ratio of communication overhead to computation carried out in parallel also decreases. This suggests that for certain classes of polynomial systems (complex function evaluation and large Jacobian matrix), the fine-grained version can perform substantially better than the serial version. In this case a mixed strategy can be employed. The coarsegrained algorithm can be used until there are no paths remaining to be tracked. Then the fine-grained algorithm can be used to finish the tracking of the uncompleted paths.
Cloud Computing
Published in Vivek Kale, Digital Transformation of Enterprise Architecture, 2019
A drawback of virtualization is the fact that the operation of the abstraction layer itself requires resources. Modern virtualization techniques, however, are so sophisticated that this overhead is not too significant: Due to the particularly effective interaction of current multicore systems with virtualization technology, this performance loss plays only a minor role in today’s systems. In view of possible savings and the quality benefits perceived by the customers, the use of virtualization pays off in nearly all cases.
GWMA: the parallel implementation of woodpecker mating algorithm on the GPU
Published in Journal of the Chinese Institute of Engineers, 2022
Jianhu Gong, Morteza Karimzadeh Parizi
The other measure that is applied to compare the sequential and parallel algorithms is the execution time. This metric, which is related to the structure and number of algorithm procedures and the problem type, is applied in the objective function. In Figure 6, execution times of 8 test problems are reported in 2000 iterations in which the problem dimension is constant and equal to 30. As reported in this figure, when the agent population is small (e.g. 30 or 100 items), no remarkable discrepancy is seen in the execution time of CWMA and GWMA. However, the diagrams of time required by both algorithms are very smooth and close to each other in lower populations, although they gradually get slopes and far from each other by increasing the population size. In other words, GWMA does not represent a high performance on a low level of tasks. In comparison, the execution time of GWMA becomes Strongly less than that of CWMA when the number of agents grows. The cause for these improvements is that the only instruction on the GPU architecture performs in a big block of data, and all data are exerted in the identical task. Undoubtedly, parallel processing in data blocks in the same period of time lowers the spending time and overhead of the computations. In addition, the search agents of the GWMA can be executed on multiple cores at the same time, and the running time will definitely be lower than using the CPU.
A parallel computing framework for solving user equilibrium problem on computer clusters
Published in Transportmetrica A: Transport Science, 2020
Xinyuan Chen, Zhiyuan Liu, Inhi Kim
The maximum speedup values of these large-scale instances vary from 45.44 to 81.09. The larger the network size (the number of OD pairs, trip origins, and links), the higher the speedup could be achieved. This is because the larger number of links, trip origins, and OD pairs allows more balanced computing resource allocation. Further, the effect of communication network overhead also become relatively smaller compared to computing overhead. In contrast, for smaller networks, the computing overhead reduces which enhances the communication overhead, due to which the idle time of unused threads could be significant. In summary, the larger network tends to have a potentially high speedup rate. The higher demand or a large network could introduce a more balanced computing task assignment across nodes and links (at different CPU cores), which leads to a potentially higher speedup rate. The average execution time of scheduling events could dramatically increase with more frequent interactions. As a result, the overall communication overhead consumes more computation resources.
Improving flexibility in cloud computing using optimal multipurpose particle swarm algorithm with auction rules
Published in Journal of Experimental & Theoretical Artificial Intelligence, 2022
Seyed Ebrahim Dashti, Mohammad Zolghadri, Fatemeh Moayedi
Moving virtual machines creates additional overhead and reduces overall efficiency. In (Xiao et al., 2019) researchers addressed the issue of reducing energy consumption while maintaining the quality of service as a multi-objective problem that tried to reduce energy consumption, ensure the quality of service and reduce the number of migrations with Double Thresholds and Ant-Colony algorithms. (Zolfaghari et al., 2021) Presents a meta-heuristic method that firstly classifies virtual machines using support vector machine. In the next step, using modified minimisation of migration, virtual machines are selected and then the particle swarm optimisation algorithm is used to place virtual machines. The main purpose of this research is to optimise energy consumption.