Research And Implementation Of Datamining Algorithm Based On Multi-core And Multi Graphics Processing Unit

In digital era, with the rapid development of network and computertechnology, all walks of life have accumulated large amounts of data. Thereforethe data mining technology especially association rules plays an important roletoday. Association rules data mining algorithms based on single core architectureare already very mature in the past. In recent years, multi-core architecturecomputing equipment have received great popularity with the rapid developmentof the hardware system, the graphics processing unit (GPU) significantlyenhanced operational capability. As a result, the hybrid multi-core CPU+GPUarchitecture for parallel data mining operation has become a new trend.Apriori algorithm is one of the most representative association rule miningalgorithm in related research methods. The traditional Apriori algorithm willcause computation time exponentially increasing when the input data get huge.As a result, the efficiency of the algorithm is very low. In addition, due to thedifferences of hardware structure, if we directly transplant the traditional singlecore architecture Apriori algorithm into the Multi-core and hybrid architecture,there will not be able to utilize hardware resources effectively making thealgorithm computation time fall further and can’t reach the goal of acceleration.Therefore, in order to improve the deficiency of Apriori algorithm and can makefull use of multi-core CPU+GPU hybrid architecture of computational resources,the research of improving the Apriori algorithm based on the Multi-core and GPUhybrid architecture is the mainly purpose and meaning of this paper.The Apriori algorithm can be mainly divided into two parts, respectively, toconfirm the candidate projects and high frequency items are combined into the candidate projects next order. To narrow the scope of the judgment and toimprove the efficiency of the merger, in this paper we propose the sorting loworder frequent set ordering strategy, thereby reducing the redundant candidateproject generated frequent itemsets, In order to achieve the goal of load balancingfor the multi-core architecture, we propose the block divide, quick merge strategybased on frequent sets utilizing reduced consolidation scope to accelerate theeffect of algorithm for different ethnic groups based on accordance with differentprojects.At the same time, in order to make full use of GPU multi-core quantity andhigh floating point arithmetic resources, we use the dynamic GPU parallelprocessing allocation mechanism according to the number of the sorted sets todetermine the number of GPU thread，then reduce the time of support numbercalculation and lower the number of checking and comparison times. As a result,it will accelerate the operation time and improve the operation efficiency, thenreturns the calculate result to CPU for the next step of operation.To test and verify the validity and effectiveness of algorithm, the paper usingthe virtual and real database platform as the simulation experimental data, designdifferent environmental parameters for the experiment. The experimental resultsshow that the algorithm runs a good performance in the multi-core platform andmulti-core and multiple graphics processing unit hybrid platform, not onlyguarantee efficiency, but also mining frequent itemsets of association rulesaccurately at the same time. Meet the high efficiency data mining needs in themulti-core and heterogeneous platform. Verified the effectiveness and feasibilityof the data mining algorithm based on multi-core and multi graphics processingunit.