Presentation July 22nd
Download as PPT, PDF•0 likes•192 views
The document proposes a cooperative caching scheme to improve traditional page fault handling in cluster computing. It aims to reduce the number and time of disk accesses during page faults by caching swapped out pages across nodes. A simulation was conducted to evaluate the feasibility of the approach, showing time savings by retrieving pages from cooperative caches over accessing local disks. The approach divides physical memory into normal and cooperative cache parts. On a page fault, it first checks local memory and cooperative caches before loading from disk.
![Yujin Tang, Hidenori Nakazato, Yoshiyori Urano Graduate School of Global Information and Telecommunication Studies Waseda University, Japan [email_address] , [email_address] , [email_address]](https://image.slidesharecdn.com/pagefaulthandlingusingcooperativecacheinhigh-090722021756-phpapp01/85/Presentation-July-22nd-1-320.jpg)
Presentation July 22nd
- 1. Yujin Tang, Hidenori Nakazato, Yoshiyori Urano Graduate School of Global Information and Telecommunication Studies Waseda University, Japan [email_address] , [email_address] , [email_address]
- 2. Introduction Characteristics of Cluster Computing About Traditional Page Fault Handling What We Did
- 3. Related Work Page Fault Handling Improvement To reduce the number of disk access (E.g., pre-mapping strategy) To reduce the time of disk access (E.g., pre-fetching strategy) Cooperative Caching Adoption To narrow the gap between memory access speed and disk access speed Different cache replacement algorithms for different application scenarios
- 4. Page Fault Handling Using Cooperative Cache Feasibility Verification Simulation Environment Item Value Operating System Solaris 10 File System UFS (synchronous write) TCP Window Size 49640 Physical Memory Size 32 GB Swap Size 16 GB CPU 2.3 GHz * 8 NIC 1 Gbps Server Load 210 queries/second
- 5. Page Fault Handling Using Cooperative Cache Feasibility Verification Simulation Result (time cost via network/time cost via disk, in μ s) Trial Page Size 4 KB 8 KB 4 MB 8 MB #1. 186/11086 378/14826 63291/3201783 95077/6355409 #2. 176/9504 361/13239 60608/3208355 99521/6358007 #3. 223/13872 369/15665 52058/3179126 115579/6433441 #4. 189/10975 339/17466 48538/3195626 97273/6463723 #5. 240/11558 363/18781 49518/3243933 111453/6527881
- 6. Page Fault Handling Using Cooperative Cache Page Fault Handling Illustration Each node maintains a set of cooperative nodes P , whose size (denoted as #P) is configurable. Kernel reserves a set of page frames T , whose size is k*#P. ( k =1, 2, …) Physical memory is divided into two parts: NU (normal use, including all free page frames) and CC (cooperative cache), their sizes are dynamically self-adjusted based on memory utilization. A requested missing page is denoted as R and a swapped out page as S.
- 7. Page Fault Handling Using Cooperative Cache Page Fault Handling Illustration Page Fault Disk CC on P S selected … Frame K 0xaaaa1000 ~ 0xaaaa1fff Frame K+1 0xaaaa2000 ~ 0xaaaa2fff … Frame N 0xbcdef000 ~ 0xbcdeffff … Physical Memory … Page T 0xabba1000 ~ 0xabba1fff Page T+1 0xabba2000 ~ 0xabba2fff … Page M 0xabdef000 ~ 0xabdeffff … Virtual Memory
- 8. Dynamic Physical Memory Distribution Page Fault Handling Using Cooperative Cache Page Fault Occurs on R Free frames in NU ? Page frames in CC ? No Select S from NU No Is S dirty? Load R into the available frame Yes Remove a page from CC Yes Write S back to disk Yes Transmit S to CC on P No
- 9. Dynamic Physical Memory Distribution Page Fault Handling Using Cooperative Cache Receive S from other node Is S in CC ? Free frames in NU ? No No Update S in CC Yes Load S into an free page frame Yes Write S back to disk Replace a frame for S in CC Page frames in CC ? Yes No
- 10. Conclusion High Page Fault Rate in Cluster Computing Inefficiency of Traditional Page Fault Handling Page Fault Handling Using Cooperative Caching Scheme
- 11. Future Work Cooperative Nodes Selection Cache Replacement Algorithm Process Migration Consideration
- 12. Q & A