1. Insertion Policy Selection Using Decision Tree Analysis Samira Khan, Daniel A. Jiménez University of Texas at San Antonio

2. Motivation  L1 and L2 filters the cache access  Last Level Cache (LLC) does not have much temporal locality  Large fraction of blocks brought to cache are never accessed again (zero reuse lines).  For SPEC CPU 2006 benchmarks, on average 60.18% lines are never accessed again while they are in the LLC

3. Motivation  No cache bursts in LLC  Only small portion of hits occur near the MRU position

4. Goal  Get rid of zero reuse lines as early as possible  Keep lines in cache for sufficient time to get the first hit  Minimal change to LRU policy  Use as little space as possible

5. Insertion Position Selection  Find the optimal insertion position  Zero reuse lines will get evicted earlier  Most of the non zero reuse lines should be in cache before their first hit  This will get rid of zero reuse lines and make space for useful lines  Use Decision Tree Analysis via set dueling to find the position  This allows choosing among the insertion positions to set duel

6. Set dueling between middle and MRU pos Set dueling between LRU and middle pos Set dueling between nearMRU and MRU pos Set dueling between nearLRU and middle pos Insert pos LRU Insert pos nearLRU Insert pos middle Insert pos nearMRU Insert pos MRU LRU posmiddle pos MRU pos nearLRU pos nearMRU pos middle pos winner MRU pos winner Middle pos winnerLRU pos winner nearMRU pos winner MRU pos winner nearLRU pos winner Middle pos winner For 400.perlbench 66.67% lines brought to cache are never accessed again and 73.03% hits occur in between MRU and middle position

7. Adaptive Multi Set Dueling  Current multi set dueling  Have one leader set for each insertion policy  Partial follower sets duplicate the winner set policy  Each policy set duel in a tournament manner  Not scalable  Leader sets performing the looser policies hurt performance  Adaptive multi set dueling  Leader set adaptively chooses the policy  No need for partial follower set  Scalable

8. Result

9. Space Overhead ParameterStorageTotal Storage LRU overhead per line4 bits1024*16*4 = 8 KB Set type per set2 bits1024 * 2 = 2048 bits Two counters (psel1 & psel2)Each 10 bits20 bits One counter (switched)1 bit Total8 KB + 2069 bits Space overhead for a 1MB 16 way set associative LLC

10. Conclusion  Insertion Position Selection using Decision Tree Analysis  Requires minimal change to LRU  Needs only 2069 bits extra space  Chooses the best insertion position adaptively  Gets rid of zero reuse lines without any storage hungry predictor  Makes multi set dueling scalable

11. Questions

12. Zero Reuse Lines in SPEC CPU 2006

13. psel ab psel cd psel ef psel gh psel 1 psel 2 psel 1 papa pbpb φ ab pcpc pdpd pepe pfpf pgpg phph φ cd φ ef φ gh pbpb papa pαpα +1 +1 +1, if p b wins -1, if p a wins All sets in LLC Leader sets in adaptive multi set dueling scheme Leader sets in current multi set dueling scheme Adaptive Multi Set Dueling

14. Result MRU nearMRU middle nearLRU LRU psel 2 psel 1 s