1. Compressed Memory Hierarchy Dongrui SHE Jianhua HUI

2. The research paper:  A compressed memory hierarchy using an indirect index cache. By Erik G. Hallnor and Steven K. Reinhardt Advanced Computer Architecture Laboratory EECS Department University of Michigan

3. Outline  Introduction  Memory eXpansion Technology  Cache-compression  IIC & IIC-C  Evaluation  Summary

4. Introduction Memory capacity and Memory bandwidth  The amount of cache cannot be increased without bound;  Scarce resource: memory bandwidth;

5. Application of data compression  First, adding a compressed main memory system (Memory Expansion Technology, MXT)  Second, Storing compressed data in the cache, then data be transmitted in compressed form between main memory and cache

6. A key challenge  Management of variable-sized data blocks: 128-byte Block After compression, 58 bytes unused

7. Outline  Introduction  Memory eXpansion Technology(MXT)  Cache-compression  IIC & IIC-C  Evaluation  Summary

8. Memory eXpansion Technology  A server class system with hardware compressed main memory.  Using LZSS compression algorithm. For most applications, two to one compression (2:l).  Hardware compression of memory has a negligible performance penalty.

9. Hardware organization  Sector translation table Each entry has 4 physical addr that each points to a 256B sector.

10. Outline  Introduction  Memory eXpansion Technology(MXT)  Cache-compression  IIC & IIC-C  Evaluation  Summary

11. Cache compression  Most designs for power savings, using more conventional cache structures: unused storage benefits only by not consuming power.  To use the space freed by compression, new cache structure is needed.

12. Outline  Introduction  Memory eXpansion Technology(MXT)  Cache-compression  IIC & IIC-C  Evaluation  Summary

13. Conventional Cache Structure  Tag associated statically with a block  When data is compressed

14. Solution: Indirect Index Cache  A tag entry not associated with a particular data block  A tag entry contains a pointer to data block

15. IIC structure  The cache can be fully associative

16. Extend IIC to compressed data  Tag contains multiple pointers to smaller data blocks

17.  Software-managed  Blocks grouped into prioritized pools based on frequency  Victim is chosen from lowest-priority non- empty pool Generational Replacement

18. Additional Cost  Compression/decompression engine  More space for the tag entries  Extra resource for replacement algorithm  Area is roughly 13% larger

19. Outline  Introduction  Memory eXpansion Technology(MXT)  Cache-compression  IIC & IIC-C  Evaluation  Summary

20. Evaluation Method: SPEC CPU2000 Benchmarks:  Main memory: 150 cycle latency, bus width 32, with MXT  L1: 1 cycle latency, split 16KB, 4-way, 64B block size  L2:12 cycle latency, unified 256KB, 8-way,128B block size  L3:26 cycle latency, unified 1MB,8- way,128B block size, with IIC-C

21. Evaluation  lsd Over 50% gain with only 10% area overhead

22. Evaluation

23. Summary Advantages:  Increase Effective Capacity & Bandwidth;  Power Saving From Less Memory Access Drawbacks:  Increase Hardware Complexity  Power Consumption of Additional Hardware

24. Future work  Overall power consumption study  Use it in embedded system

25. END Thank you ! Question time.