软件学报
軟件學報
연건학보
JOURNAL OF SOFTWARE
2014年
8期
1844-1857
,共14页
徐传福%车永刚%王正华%彭宇行
徐傳福%車永剛%王正華%彭宇行
서전복%차영강%왕정화%팽우행
计算机体系结构%分布模拟%负载均衡%可扩展
計算機體繫結構%分佈模擬%負載均衡%可擴展
계산궤체계결구%분포모의%부재균형%가확전
computer architecture%distributed simulation%load balance%scalability
分布式并行模拟是提高体系结构模拟速度的有效技术手段之一.首先,建立了分布式并行模拟的通用性能分析模型,并对典型系统的并行加速比、并行效率等性质进行了理论分析,得出了一些有用的结论.在此基础上,提出了均衡可扩展分布式并行模拟方法SEDSim(scalable and evenly distributed simulation).SEDSim针对模拟节点负载不均衡问题,提出了开销模型指导的指令区间均衡分割和分配策略CoMEPA(cost model guided evenly partition and allocation);针对分布式并行模拟与非连续、任意数量抽样模拟区间的高效集成,提出了基于最小等价距离(minimum equivalent cost,简称MinEC)的指令区间分配策略MinEC.基于sim-outorder实现了SEDSim,采用SPEC CPU2000中的部分程序对其速度和精度进行了测试,理论分析和测试结果均表明了SEDSim的优势:相对于常用的方法或策略, CoMEPA和MinEC分别能够获得多达约1.6倍和1.4倍的性能提升.
分佈式併行模擬是提高體繫結構模擬速度的有效技術手段之一.首先,建立瞭分佈式併行模擬的通用性能分析模型,併對典型繫統的併行加速比、併行效率等性質進行瞭理論分析,得齣瞭一些有用的結論.在此基礎上,提齣瞭均衡可擴展分佈式併行模擬方法SEDSim(scalable and evenly distributed simulation).SEDSim針對模擬節點負載不均衡問題,提齣瞭開銷模型指導的指令區間均衡分割和分配策略CoMEPA(cost model guided evenly partition and allocation);針對分佈式併行模擬與非連續、任意數量抽樣模擬區間的高效集成,提齣瞭基于最小等價距離(minimum equivalent cost,簡稱MinEC)的指令區間分配策略MinEC.基于sim-outorder實現瞭SEDSim,採用SPEC CPU2000中的部分程序對其速度和精度進行瞭測試,理論分析和測試結果均錶明瞭SEDSim的優勢:相對于常用的方法或策略, CoMEPA和MinEC分彆能夠穫得多達約1.6倍和1.4倍的性能提升.
분포식병행모의시제고체계결구모의속도적유효기술수단지일.수선,건립료분포식병행모의적통용성능분석모형,병대전형계통적병행가속비、병행효솔등성질진행료이론분석,득출료일사유용적결론.재차기출상,제출료균형가확전분포식병행모의방법SEDSim(scalable and evenly distributed simulation).SEDSim침대모의절점부재불균형문제,제출료개소모형지도적지령구간균형분할화분배책략CoMEPA(cost model guided evenly partition and allocation);침대분포식병행모의여비련속、임의수량추양모의구간적고효집성,제출료기우최소등개거리(minimum equivalent cost,간칭MinEC)적지령구간분배책략MinEC.기우sim-outorder실현료SEDSim,채용SPEC CPU2000중적부분정서대기속도화정도진행료측시,이론분석화측시결과균표명료SEDSim적우세:상대우상용적방법혹책략, CoMEPA화MinEC분별능구획득다체약1.6배화1.4배적성능제승.
Distributed simulation is an effective method to improve simulation speed for computer architecture. In this paper, a general performance model for distributed simulation is established and then some typical distributed simulation systems are analyzed based on the model. The analysis results in some important conclusions about parallel speedup and parallel efficiency for distributed simulation. Next, a scalable and evenly distributed simulation (SEDSim) approach is presented. SEDSim adopts a cost model guided even partition and allocation (CoMEPA) policy for benchmark program instructions to enhance load-balance among parallel simulation nodes. An allocation policy based on minimum equivalent cost (MinEC) is also designed to efficiently integrate arbitrary number of discrete sampling intervals in SEDSim. The study implementes SEDSim based on sim-outorder and evaluates its speed and accuracy using Benchmark programs from SPEC CPU2000. Both theoretical analysis and testing results validate some advantages of SEDSim approach. Compared with existing methods, CoMEPA and MinEC can achieve a speedup of about 1.6 and 1.4 respectively.