1. Performance of Snooping Protocols Kay Jr-Hui Jeng

2. Outline Snooping protocols Snooping protocols Simulation results Simulation results Comparison of performance Comparison of performance Conclusion Conclusion

3. Snooping protocols a protocol for maintaining cache coherency in symmetric multiprocessing environments. In a snooping system, all caches on the bus snoop the bus to determine if they have a copy of the block of data that is requested on the bus. Every cache has a copy of the sharing status of every block of physical memory it has. Multiple copies of a document in a multiprocessing environment typically can be read without any coherence problems; however, a processor must have exclusive access to the bus in order to write. a protocol for maintaining cache coherency in symmetric multiprocessing environments. In a snooping system, all caches on the bus snoop the bus to determine if they have a copy of the block of data that is requested on the bus. Every cache has a copy of the sharing status of every block of physical memory it has. Multiple copies of a document in a multiprocessing environment typically can be read without any coherence problems; however, a processor must have exclusive access to the bus in order to write.

4. Comparisons of Write-invalidate & Write-update Write-invalidateWrite-update Writing Writing processor forces all others to invalidate their copies Writing processor forces all others to update their copies Advantage Less bus traffic Other processors can get the data faster disadvantage Update the data to cache slower Higher bus traffic

5. Comparisons of Write-back & Write-through Write-backWrite-through Writing The memory is updated only when the block in the cache is being replace The memory is updated every the cache is updated Advantage Less bus traffic Memory can get the new data faster Every write is observable-> Simple disadvantage Update the data to memory slower Higher bus traffic

6. Types of Snooping Protocols Write-Invalidate and Write-Through Write-Invalidate and Write-Through Write-Invalidate and Write-Back Write-Invalidate and Write-Back Write-Once Write-Once Write-Update and Partial Write-Through (Firefly) Write-Update and Partial Write-Through (Firefly) Write-Update and Write-Back (Dragon) Write-Update and Write-Back (Dragon)

7. Write-Invalidate and Write-Through The memory is always consistent with the most recently updated cache copy The memory is always consistent with the most recently updated cache copy Multiple processors can read block copies from main memory safely until one processor updates its copy. At this time, all cache copies are invalidated and the memory is updated to remain consistent. Multiple processors can read block copies from main memory safely until one processor updates its copy. At this time, all cache copies are invalidated and the memory is updated to remain consistent.

8. Write-Invalidate and Write-Through VALID INV Bus Write-Miss, Bus Write-Hit Write-Hit, Read-Hit Write-Miss, Read-Miss Read-Miss

9. Synapse It belongs to write-invalidate & write-back protocol It belongs to write-invalidate & write-back protocol A multiprocessor for fault-tolerant transaction processing A multiprocessor for fault-tolerant transaction processing It has two system buses, the added bandwidth of the extra bus allows the system to be expanded to more processors (max 28) It has two system buses, the added bandwidth of the extra bus allows the system to be expanded to more processors (max 28) The including single-bit tag with each cache block in main memory =>whether main memory is to respond to a miss on that block The including single-bit tag with each cache block in main memory =>whether main memory is to respond to a miss on that block

10. Synapse

11. Write-Once Write-invalidate protocol (Goodman) Write-invalidate protocol (Goodman) Designed for single-board using Multi-bus Designed for single-board using Multi-bus Combination of Write-through and Write-back Combination of Write-through and Write-back

12. Write-Once

13. Write-Update and Partial Write-Through (Firefly) An update to one cache is written to memory at the same time; Broadcast to other caches sharing the updated block An update to one cache is written to memory at the same time; Broadcast to other caches sharing the updated block These caches snoop on the bus and perform updates to their local copies. These caches snoop on the bus and perform updates to their local copies. There is also a special bus line, which is used to detect sharing There is also a special bus line, which is used to detect sharing Multiple writers are permitted-the data for each write to a shared block are transmitted to each and to the backing store =>The protocol never causes an invalidation Multiple writers are permitted-the data for each write to a shared block are transmitted to each and to the backing store =>The protocol never causes an invalidation

14. Write-Update and Partial Write- Through (Firefly)

15. Write-Update and Write-Back (Dragon) It similar to write-update and partial write through It similar to write-update and partial write through Memory updates are done only when the block is being replaced Memory updates are done only when the block is being replaced Writes to shared blocks are not immediately sent to main memory, only to other caches that have a copy of the block Writes to shared blocks are not immediately sent to main memory, only to other caches that have a copy of the block

16. Write-Update and Write-Back (Dragon)

17. Experiment Results (1)

18. Experiment Results (2)

19. Experiment Results (3)

20. Experiment Results (4)

21. Experiment Results (5)

22. Result analysis (1) F-7: Dragon and Firefly protocols are identical in the handling of private blocks. The performance of write-once is dependent on the trade-off between single word writes and the reduction in the write-back. The performance of Synapse is below the others as a result of the additional overhead of treating write hits on unmodified block as write misses. F-7: Dragon and Firefly protocols are identical in the handling of private blocks. The performance of write-once is dependent on the trade-off between single word writes and the reduction in the write-back. The performance of Synapse is below the others as a result of the additional overhead of treating write hits on unmodified block as write misses.

23. Result analysis (2) Comparison of F-8~F-10 with F-7 indicates the impact of handling shared blocks efficiently Comparison of F-8~F-10 with F-7 indicates the impact of handling shared blocks efficiently F-8, F-11 and F-14 (share block=16), which demonstrate that the distributed write approach of Dragon and Firefly has the best performance in the handling of shared data. F-8, F-11 and F-14 (share block=16), which demonstrate that the distributed write approach of Dragon and Firefly has the best performance in the handling of shared data. F-11, F-12 and F-13: Because no invalidations, the performance of Dragon and Firefly decreases as the average actual sharing decreases and the number of shared blocks increases. F-11, F-12 and F-13: Because no invalidations, the performance of Dragon and Firefly decreases as the average actual sharing decreases and the number of shared blocks increases.

24. Result analysis (3) The performance of the Dragon exceeds that of the Firefly at levels of high sharing(F-8 and F-9) because the Firefly must send distributed writes to global memory while the Dragon sends them to the caches only. The performance of the Dragon exceeds that of the Firefly at levels of high sharing(F-8 and F-9) because the Firefly must send distributed writes to global memory while the Dragon sends them to the caches only. The performance of write-once is lower than above protocols as a result of the added overhead of updating memory each time a Dirty block is missed in another cache. The performance of write-once is lower than above protocols as a result of the added overhead of updating memory each time a Dirty block is missed in another cache.

25. Result analysis (4) The performance of Synapse is lower, because it increased overhead of read misses on blocks that are Dirty in another cache and to the added overhead of loading new data on a write hit on an unmodified block. The performance of Synapse is lower, because it increased overhead of read misses on blocks that are Dirty in another cache and to the added overhead of loading new data on a write hit on an unmodified block. The results of Write-Invalidate and Write- Through which has the lowest performances, because blocks are not loaded into the cache on a write miss. The results of Write-Invalidate and Write- Through which has the lowest performances, because blocks are not loaded into the cache on a write miss.

26. References Hesham El-Rewini, Mostafa Abd-El-Barr. Advanced Computer Architecture and Parallel Processing, John Wiley, 2005. Hesham El-Rewini, Mostafa Abd-El-Barr. Advanced Computer Architecture and Parallel Processing, John Wiley, 2005. James Archibald and Jean-Loup Baer. Cache Coherence Protocols: Evaluation Using a Multiprocessor Simulation Model. ACM Transactions on Computer Systems, Vol. 4, No. 4, November 1986, 273-298(1986). James Archibald and Jean-Loup Baer. Cache Coherence Protocols: Evaluation Using a Multiprocessor Simulation Model. ACM Transactions on Computer Systems, Vol. 4, No. 4, November 1986, 273-298(1986). http://www.webopedia.com/TERM/S/snooping_prot ocol.html http://www.webopedia.com/TERM/S/snooping_prot ocol.html http://www.webopedia.com/TERM/S/snooping_prot ocol.html http://www.webopedia.com/TERM/S/snooping_prot ocol.html

27. References https://www.cs.tcd.ie/Jeremy.Jones/vivio/caches/writ eOnceHelp.htm https://www.cs.tcd.ie/Jeremy.Jones/vivio/caches/writ eOnceHelp.htm https://www.cs.tcd.ie/Jeremy.Jones/vivio/caches/writ eOnceHelp.htm https://www.cs.tcd.ie/Jeremy.Jones/vivio/caches/writ eOnceHelp.htm http://cs- www.bu.edu/faculty/best/crs/cs551/lectures/lecture- 12.html http://cs- www.bu.edu/faculty/best/crs/cs551/lectures/lecture- 12.html http://cs- www.bu.edu/faculty/best/crs/cs551/lectures/lecture- 12.html http://cs- www.bu.edu/faculty/best/crs/cs551/lectures/lecture- 12.html http://cs.engr.uky.edu/~manivann/Spring2005/cs570 /ch5-6.pdf http://cs.engr.uky.edu/~manivann/Spring2005/cs570 /ch5-6.pdf http://cs.engr.uky.edu/~manivann/Spring2005/cs570 /ch5-6.pdf http://cs.engr.uky.edu/~manivann/Spring2005/cs570 /ch5-6.pdf www.cs.berkeley.edu/~pattrsn/252F96/Lecture18.pdf www.cs.berkeley.edu/~pattrsn/252F96/Lecture18.pdf www.cs.berkeley.edu/~pattrsn/252F96/Lecture18.pdf http://www.eng.tau.ac.il/~arch/LECTURES/Lecture 1.pdf http://www.eng.tau.ac.il/~arch/LECTURES/Lecture 1.pdf http://www.eng.tau.ac.il/~arch/LECTURES/Lecture 1.pdf http://www.eng.tau.ac.il/~arch/LECTURES/Lecture 1.pdf

28. Thank You!