網格性能問題初探_第1頁
網格性能問題初探_第2頁
網格性能問題初探_第3頁
網格性能問題初探_第4頁
網格性能問題初探_第5頁
已閱讀5頁,還剩46頁未讀, 繼續免費閱讀

下載本文檔

版權說明:本文檔由用戶提供并上傳,收益歸屬內容提供方,若內容存在侵權,請進行舉報或認領

文檔簡介

1、網格性能問題初探網格監控和性能小組霍志剛 莫志鋒2002年12月26日內容簡介 問題的提出 網格性能模型 GloPerf MPI-G簡介網格性能是被大家遺忘的角落性能 : (Grid Computing Grids: enable communities (“virtual organizations”) to share geographically distributed resources as they pursue common goals - assuming the absence of central location, central control, omniscience

2、, existing trust relationships. Major differences to parallel systems Dynamic system of resources Large number of diverse systems Sharing of resources Transparent resource allocation 大部分應用常常只能得到5%甚至1%以下利用率 性能價格比 - - 總擁有性能價格比Total Performance/Cost of Ownership 性能無關的因素 越來越重要互操作性、可移植性、安全性、好用性、易管理性和高可用性

3、易部署性 服務協議等級(Service Level Agreement) From 課程講義 gct01.ppt問題的提出 理論上,網格環境具有“無限的”可擴展性。但是,當網格的規模擴大時,性能的惡化是任何網格服務,尤其是分布式計算服務,所難以回避的一個難題。 平臺無關的東西最終還是要運行在一定的平臺上。內容簡介 問題的提出 網格性能模型 GloPerf MPI-G簡介網格性能模型 Application Signature Model Application Intrinsic Metrics(應用本征值) description of application demands on reso

4、urces sample metrics ratio of memory references to arithmetic/logical operations, memory reference patterns, ILP, thread level parallelism, and I/O characteristics FLOPS/statement, I/O bytes/statement, bytes/message values may depend on problem parameters Application Signature trajectory of values t

5、hrough N-dimensional metric space one trajectory per taskApplication Signature ModelM1M2M3 GrADS Project System Space Metrics description of resource response to application demands sample metrics FLOPS/second, I/O bytes/second, message bytes/second values are dependent on execution platform quantif

6、y actual performance System Space Signature trajectory of values through N-dimensional metric space will vary across application executions, even on the sameresourcesSystem Space SignatureM1M2M3 GrADS ProjectPerformance Prediction Strategy Given application intrinsic behavior resource capability inf

7、ormation project application signature into system space signature, in effect predicting performance Many possible projection strategies single figure of merit (scaling in each dimension) peak MFLOPS, bandwidth, I/O rate benchmark suite measurements previous application executions (learning) GrADS P

8、rojectManaging Dynamism: The GrADS Approach GrADS (Grid Application Development Software) Funded by NSF, started 2000 Goal:Provide application development technologies that make it easy to construct and execute applications with reliable and often high performance in the constantly-changing environm

9、ent of the Grid. Major techniques to handle transparency and dynamism: Dynamic configuration to available resources (configurable object programs) Performance contracts and dynamic reconfigurationGrADS Software ArchitecturePSEConfig.objectprogramwholeprogramcompilerSource appli-cationlibrariesRealti

10、meperfmonitorDynamicoptimizerGrid runtime System(Globus)negotiationSoftware ComponentsScheduler/ServiceNegotiatorPerformance feedbackProgram Preparation SystemExecution Environment性能合約( contracts)性能合約是在一定的性能模型的基礎上,對一個網格應用的可測量的各項性能指標的約定。即,給定: set of resources, capabilities of resources, problem param

11、eters 則,該應用程序將會 達到某個指定的可測量的性能性能模型的來源 Developer knowledge of application or library behavior Compile time analysis of code Use compiler understanding of code behavior to build performance predictions Historical data from previous runs or observed behavior of current execution so far “Learn” from past

12、 experience Application Signature Model is one example From GrADS Project性能合約的生成ProgramPerformanceModelResource BrokerResourceAssignmentPerformanceContract Developer Compiler MeasurementsMDSNWS性能合約的驗證ExecutionContractMonitorReschedulingSensor DataSteerDynamic Optimizer Violation detection Fault dete

13、ction合約監控程序(Contract Monitor) 由三個用戶配置參數控制 監控時間片長度(Time quantum for “time per iteration”) 違約程度容限(% degradation in time per iteration (relative to prior average) before noting violation) 違約次數容限(Number of violations before migration)合約監控程序(續) 違約的可能原因 Competing load on CPU Computation requires more proc

14、essing power Hardware problems合約監控過程 The end user specifies several variables. These variables can be changed during runtime by contacting the application with an HTTP interface. The system will then calculate the average wall clock time per iteration for each time quantum. If the average iteration

15、in any time quantum has lower performance (by the percentage specified) than the average for all the other previous quanta, then a violation is noted.Actions Taken on Contract Violation Occurs when more than the specified number of violations have been noted New set of resources requested from the R

16、esourceSelector Checkpoints application Moves checkpoint data to the new resources along with other data needed for restart Restarts application on the new resources內容簡介 問題的提出 網格性能模型 GloPerf MPI-G簡介GloPerfGlobus Network Performance Measurement Tool GloPerf 在任意一對IP之間定期執行網絡性能測試軟件netperf的庫版本。 GloPerf測試

17、方法 TCP_STREAM測試節點間吞吐率。 TCP_RR(即Request/Response)測試延遲。 From Globus ProjectGloPerf 的測試過程 每個GloPerf后臺進程都維護著一個由MDS查詢得到的其它GloPerf后臺進程的IP地址列表。每隔5分鐘,它從中選取一個目的地址進行測試。為了減少測試沖突帶來的測量誤差/錯誤,對象的選取應該是隨機的。GloPerf的流程的偽碼while ( no termination signal ) query MDS for all gloperfds;filter list of gloperfds on version nu

18、mber and groups;build a randomized list of tests that does abandwidth test and a latency test to each filtered peer;for each element of the list perform the test;write the results in the MDS;wait n minutes;一些實測數據以下三個圖的數據是在 Globus testbed(GUSTO)上,在1999年8月到10月間,從138個GloPerf后臺進程采集的數據中提取出來的(包括3158對計算機的1

19、7629個測量數據)。這些計算機大部分位于北美,還有一些位于亞洲、歐洲和澳大利亞。 From Globus Project3內容簡介 問題的提出 網格性能模型 GloPerf MPI-G簡介并行程序編程模型 共享存儲 DSM 消息傳遞 MPIMPI簡介 Message Passing Interface由40多個組織組成的MPI Forum從1992年起共同制定。 MPICH implementation was developed and distributed by the Argonne National Laboratory (ANL) MPICH group. 可移植性好: any

20、correct program compiles & runs on a platform 無互操作性: all processes must link against same SDK E.g., MPICH and LAM versions of MPIApplicationApplicationMPI APIMPI APILAM SDKLAM protocolMPICH-P4 SDKMPICH-P4 protocolTCP/IPTCP/IPDifferent message formats, exchange sequences, etc.MPI的特點一個簡單的MPI程序#inc

21、lude “mpi.h” /* 省略了其它“常見的”C頭文件。*/int main(int argc,char *argv) int n, myid, numprocs; double mypi, pi; MPI_Init(&argc,&argv); MPI_Comm_size(MPI_COMM_WORLD,&numprocs); MPI_Comm_rank(MPI_COMM_WORLD,&myid); if( myid = 0 ) n = 99; MPI_Bcast(&n, 1, MPI_INT, 0, MPI_COMM_WORLD); MPI_Red

22、uce(&mypi, &pi, 1, MPI_DOUBLE, MPI_SUM, 0, MPI_COMM_WORLD); MPI_Finalize(); return 0;MPICH-G2: A Grid-Enabled MPI 是MPICH在基于Globus的一個實現版本。 由ANL基于其MPICH實現 將MPI和Globus提供的各種服務相結合,例如,任務啟動、認證、安全、數據轉換、 文件存取等等。 在廣域網環境下,原有程序無需任何修改,重新編譯即可運行。/mpi Golbus的層次化體系結構ApplicationsCore ServicesMet

23、acomputing Directory ServiceGRAMGlobus Security InterfaceReplica CatalogGASSGridFTPLocal ServicesLSFCondorMPINQEPBSTCPAIXLinuxUDPHigh-level Services and ToolsCactusCondor-GMPINimrod/GglobusrunPUNCHGrid StatusI/OSolarisDRM測試環境 測試平臺是三個本地結點和一個遠程結點,其中遠程結點通過2Mbps的線路從廣域網上與本地相連。每個結點運行的是由Globus1.1.4定制的INFN-

24、GRID 1.1.2。測試工具 Mpptest SMP and LAN tests mpirun -np 2 mpptest -reps 4 -size 0 50000 1000 (to get bandwidth) mpirun -np 2 mpptest -reps 4 -size 0 50 1 (to get latency) WAN tests mpirun -globusrsl mympptest.rsl RSL文件格式mympptest.rsl+(&(resourceManagerContact=janus.pr.infn.it) (count= 1) (label=subjob 0) (environment=(GLOBUS_DUROC_SUBJOB_INDEX 0) (arguments= -reps 10 -size 0 50 2 ) (directory=/home/alfieri) (executable=/home/alfieri/mpptest)(&(resour

溫馨提示

  • 1. 本站所有資源如無特殊說明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請下載最新的WinRAR軟件解壓。
  • 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請聯系上傳者。文件的所有權益歸上傳用戶所有。
  • 3. 本站RAR壓縮包中若帶圖紙,網頁內容里面會有圖紙預覽,若沒有圖紙預覽就沒有圖紙。
  • 4. 未經權益所有人同意不得將文件中的內容挪作商業或盈利用途。
  • 5. 人人文庫網僅提供信息存儲空間,僅對用戶上傳內容的表現方式做保護處理,對用戶上傳分享的文檔內容本身不做任何修改或編輯,并不能對任何下載內容負責。
  • 6. 下載文件中如有侵權或不適當內容,請與我們聯系,我們立即糾正。
  • 7. 本站不保證下載資源的準確性、安全性和完整性, 同時也不承擔用戶因使用這些下載資源對自己和他人造成任何形式的傷害或損失。

最新文檔

評論

0/150

提交評論