安徽理工大学学报（自然科学版）
安徽理工大學學報（自然科學版）
안휘리공대학학보（자연과학판）
Journal of Anhui University of Science and Technology (Natural Science)
2015年
3期
30-35
,共6页
时延Petri网%多计算单元%任务调度%torus互连网络
時延Petri網%多計算單元%任務調度%torus互連網絡
시연Petri망%다계산단원%임무조도%torus호련망락
time petri net%multi-computing units%task schedule%torus interconnection network
多核任务映射是高性能计算领域的研究热点。基于torus互连的多核架构，本文提出了行资源映射和行优化资源映射两种算法，行资源映射算法首先以torus的最左边位置作为起点映射，然后按入度值从大到小的次序依次映射其后继，每映射完一个后继就映射其相应的前驱，以此类推，映射完所有的节点。行优化资源映射首先把出度最大的点放在torus最左边位置，把入度最大的点放在torus同一行最右边位置，然后映射出度最大的点的其他后继和入度最大的点其他前驱，以此类推，映射完所有的节点。利用时延petri网软件工具包对两种映射方法进行了分析和比较，实验结果表明，基于相同的循环任务数据流图和硬件架构，行优化资源映射算法获得的吞吐量是行资源映射算法的2倍，并且每个映射块运行时间减少近似一半，行优化资源映射算法具有合理性和可行性。
多覈任務映射是高性能計算領域的研究熱點。基于torus互連的多覈架構，本文提齣瞭行資源映射和行優化資源映射兩種算法，行資源映射算法首先以torus的最左邊位置作為起點映射，然後按入度值從大到小的次序依次映射其後繼，每映射完一箇後繼就映射其相應的前驅，以此類推，映射完所有的節點。行優化資源映射首先把齣度最大的點放在torus最左邊位置，把入度最大的點放在torus同一行最右邊位置，然後映射齣度最大的點的其他後繼和入度最大的點其他前驅，以此類推，映射完所有的節點。利用時延petri網軟件工具包對兩種映射方法進行瞭分析和比較，實驗結果錶明，基于相同的循環任務數據流圖和硬件架構，行優化資源映射算法穫得的吞吐量是行資源映射算法的2倍，併且每箇映射塊運行時間減少近似一半，行優化資源映射算法具有閤理性和可行性。
다핵임무영사시고성능계산영역적연구열점。기우torus호련적다핵가구，본문제출료행자원영사화행우화자원영사량충산법，행자원영사산법수선이torus적최좌변위치작위기점영사，연후안입도치종대도소적차서의차영사기후계，매영사완일개후계취영사기상응적전구，이차유추，영사완소유적절점。행우화자원영사수선파출도최대적점방재torus최좌변위치，파입도최대적점방재torus동일행최우변위치，연후영사출도최대적점적기타후계화입도최대적점기타전구，이차유추，영사완소유적절점。이용시연petri망연건공구포대량충영사방법진행료분석화비교，실험결과표명，기우상동적순배임무수거류도화경건가구，행우화자원영사산법획득적탄토량시행자원영사산법적2배，병차매개영사괴운행시간감소근사일반，행우화자원영사산법구유합이성화가행성。
Multi-core task mapping is a hot research topic in the field of high performance computing tasks . Based on torus interconnect multi -core architecture , the RRM ( row resource mapping ) algorithm and RORM ( row optimum resource mapping ) algorithm were proposed .The mapping starting point of RRM is the leftmost position in torus at first .Then its successors are mapped in the order of indegree values .Corresponding prede-cessors are mapped after mapping a successor and so on .Finally, all of the nodes are mapped completely .The maximum outdegree node is the mapping starting point of RORM , and it is mapped onto the leftmost position of torus at first , and the maximum indegree node is mapped onto the rightmost position of torus .Then the other suc-cessors of the maximum outdegree node and the other predecessors of the maximum indegree node are mapped onto torus.Finally, all of the nodes are mapped completely .Two mapping methods were analyzed and compared by time petri net software tool kits .Based on the same cyclic data flow graph and hardware architecture , the re-sults of experiment showed that the throughput of RORM is 2 times of RRM.Compared with RRM, the running time of each mapping blocks by RORM reduced approximately 50%.RORM is rational and feasible .
