1. 1 Virtual Private Caches ISCA’07 Kyle J. Nesbit, James Laudon, James E. Smith Presenter: Yan Li

2. 2 CMP-based System Chip-level Multiprocessor  multiple processor cores are implemented into a single chip  Multithreading support Intel Core 2 Duo E6750

3. 3 CMP-based System (2) Resource sharing  Cache capacity/bandwidth, main memory…… Pros: Higher resource utilization Cons: Inter-thread interference  Unpredictable performance / no QoS! Many applications running on CMP-based systems require Quality of Service

4. 4 Quality of Service QoS are required by many applications:  Soft real-time applications video games  Find-grain parallel applications Scheduling & synchronization  Server consolidation Hosting services QoS objectives in CMP-based system  provide an upper bound on thread execution time regardless of other thread activity

5. 5 Outline Introduction QoS Framework Virtual Private Cache - VPC Arbiter Virtual Private Cache - Capacity Manager Performance Evaluation Conclusions

6. 6 Overview of VPM Virtual Private Machine: A set of allocated hardware resources  Processors, bandwidth, memory spaces… Each thread is allocated a share of hardware resource based on policies  Applications & system software Hardware mechanism enforces allocated resources

7. 7 System hardware VPM

8. 8 Objectives of VPM Performance Isolation  thread performance is as good as on real private machine having same resources Dynamic distribution of excess resources  Unallocated resources  Allocated but not used resources

9. 9 Virtual Private Cache Microarchitecture-level mechanism Main components  VPC Arbiter: tag & data array bandwidth sharing  VPC Capacity Manager: cache capacity sharing Advantages  Performance isolation  Improved utilization

10. 10 Outline Introduction QoS Framework Virtual Private Cache - VPC Arbiter Virtual Private Cache - Capacity Manager Performance Evaluation Conclusions

11. 11 VPC Arbiter - Implementation(1) Each data & tag array has an arbiter Each arbiter has  FIFO buffer for each thread:  1 clock register R.clk: determine arrival time  R.Li & R.Si for thread i: virtual service/start time

12. 12 VPC Arbiter - Implementation(2) R.Li: virtual service time of a request from thread i   L: latency of shared cache; : thread i’s fraction of resources R.Si: virtual start time of the next request of thread i  Time that the resource is available for the next request of thread i

13. 13 Fair Queuing Scheduling Request Arrival:  Arbiter Calculation of virtual finish time:  Arbiter Selection:  select the request with the earliest Fi 

14. 14 Arbiter Fairness Policy Excess bandwidth is distributed to threads that has received the least excess bandwidth in the past

15. 15 Outline Introduction QoS Framework Virtual Private Cache - VPC Arbiter Virtual Private Cache - Capacity Manager Performance Evaluation Conclusions

16. 16 Implementation Set associative replacement policy Each thread receives  same number of sets as the shared cache  at least Replacement policy  LRU line owned by thread i, such that thread i owns more than ways  LRU line owned by the thread that requesting the replacement

17. 17 Outline Introduction QoS Framework Virtual Private Cache - VPC Arbiter Virtual Private Cache - Capacity Manager Performance Evaluation Conclusions

18. 18 Experiment Setup Two microbenchmarks to stress performance isolation feature  Loads: load operations with continuous read hits  Stores: store operations with continuous write hits SPEC CPU2000 benchmark suite QoS performance metrics  IPC  Data array utilization

19. 19 Other Arbiter Read over Write  Prioritize read over write Read over Write First Come First Service  Prioritize read over write  Prioritize oldest requests Round Robin  Interleave requests uniformly and consistently

20. 20 Microbenchmark

21. 21 SPEC

22. 22 Conclusions VPC: hardware mechanism of VPM QoS framework  VPC arbiter & capacity manager VPC can achieve global QoS objectives Issues:  Local QoS objectives assumes performance monotonicity

23. 23 Thank You! & Questions?