Write-Rationing Garbage Collection for Hybrid Memories

Transcript

1. Write-Rationing Garbage Collection for Hybrid Memories Shoaib Akram, Jennifer B. Sartor, Kathryn S. Mckinely, Lieven Eeckhout Ghent University, Belgium Shoaib.Akram@UGent.be
2. DRAM as main memory is facing multiple challenges Energy Scalability Cost Reliability
3. Opportunity for new memory technologies to replace DRAM Source: https://www.nextplatform.com/2015/07/29/scaling-the-growing-system-memory-hierarchy/
4. PCM cells have limited write endurance, shortening its lifetime Current (Temperature) Time Read Reset to amorphous Set to crystalline 610°C 350°C
5. Speed ++ Lifetime ++ Energy Density DRAM Speed Lifetime Energy ++ Density ++ Hybrid memory is the best of DRAM and PCM PCM
6. DRAM PCM Future of main memory: some DRAM and lots of PCM
7. DRAM PCM Challenge: limit # writes to PCM Solution: Use DRAM for frequently written data
8. Challenge: limit # writes to PCM Solution: Use DRAM for frequently written data
9. Memory automatically reclaimed for reuse More than just reclaim, stuff better organized Garbage collection: key advantage of using a managed language
10. Memory automatically reclaimed for reuse More than just reclaim, stuff better organized Our Contribution Write-Rationing Garbage Collection Ration (verb) To restrict the consumption of (a commodity, food, etc.) dictionary.com
11. Modern collectors segregate young and old objects •  Nursery objects die quickly •  When nursery is full – Traced for live objects – Survivors copied to mature space Nursery (young) Mature (old)
12. Young objects incur 70% of all writes to Java applications 0 25 50 75 100 % of writes Nursery 70
13. Old objects incur 30% of all writes to Java applications 0 25 50 75 100 % of writes Mature Nursery 30 70
14. Top 10% most-written to old objects receive 29% of writes 0 25 50 75 100 % of writes Mature Nursery Top 10% 30 70
15. Top 2% most-written to old objects receive 24% of writes 0 25 50 75 100 % of writes Mature Nursery Top 10% 30 70 24 Top 2%
16. The two enemies of PCM lifetime: nursery and writers Enemy no. 1: Nursery Enemy no. 2: Writers Nursery (young) Mature (old)
17. Our proposed collectors protect PCM from nursery and writers Kingsguard-Nursery (KG-N) Kingsguard-Writers (KG-W)
18. Kingsguard-Nursery: protect PCM from nursery nursery mature meta data large mutator mutator runtime GC DRAM
19. Kingsguard-Nursery: protect PCM from nursery nursery mature meta data large mutator mutator runtime GC DRAM PCM JVM OS PCM DRAM
20. Kingsguard-Writers: protect PCM from nursery and writers nursery mature DRAM observer PCM mature meta data large mutator mutator large How to monitor writes to observers?
21. Nursery (young) Mature (old) Roots Live object Write barrier: extra work on each pointer update outside nursery •  Don’t want to scan all objects in mature •  Remember mature to young pointers in a set Remembered set
22. Kingsguard-Writers uses write barriers for monitoring writes Primitive Fields Reference Fields Object Meta-Data Write Intensity if (dst_object outside nursery) add src_object to remembered set update_write_intensity(src_object) if (src_object outside nursery) update_write_intensity(src_object) reference_barrier (src_object, dst_object) primi<ve_barrier (src_object)
23. Kingsguard-Writers does additional optimizations to protect PCM 1.  Placing PCM meta-data in DRAM 2.  Allocating large objects in nursery
24. Kingsguard-Writers optimization 2: Placing PCM meta-data in DRAM Mature (PCM) A B C D . . . Mature graph traversal PCM LLC . A Writes to PCM
25. Kingsguard-Writers optimization 2: Placing PCM meta-data in DRAM Mature (PCM)
26. Kingsguard-Writers optimization 2: Placing PCM meta-data in DRAM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mature (PCM) Meta (DRAM)
27. Kingsguard-Writers optimization 2: Placing PCM meta-data in DRAM Mature (PCM) Meta (DRAM) ........... ........... ........... ...........
28. Kingsguard-Writers optimization 2: Placing PCM meta-data in DRAM Mature (PCM) Meta (DRAM) ........... ........... ........... ........... P start_address + index_of_object
29. Kingsguard-Writers optimization 2: Meta-data in DRAM overhead Mature (PCM) Meta (DRAM) ........... ........... ........... ........... Smallest object = 4 B 25% overhead Common case > 16 B 6.25% overhead For objects smaller than 16 B : use header byte On allocation, set a bit in header if object size < 16 B P start_address index_of_object +
30. Kingsguard-Writers: protect PCM from nursery and writers nursery mature DRAM observer PCM mature meta data large mutator mutator large runtime meta data
31. Kingsguard-Writers optimization : Allocating large objects in nursery Nursery (DRAM) ½ of remaining nursery Large (PCM) if (alloc_rate_large >> alloc_rate_nursery) then trigger_large_opt()
32. Kingsguard-Writers: protect PCM from nursery and writers nursery mature DRAM observer PCM mature meta data large mutator mutator large runtime meta data
33. KG-N & KG-W summary KG-N KG-W •  Hybrid memory aware JVM •  Nursery in DRAM •  Observers •  Mature DRAM and mature PCM •  Large object opt •  Meta-data opt
34. Experimental Setup ² Fast architecture simulator ² Helped us gain insight Example: track origin of writes ² Easy to modify ² Framework to control non-determinism
35. Benchmark Characteristics Benchmark Function Heap Size in MB Allocation in MB % Nursery survival Xalan XML parser 108 980 17 Pmd Code analysis 98 364 23 Lusearch Text search 68 4294 4 Bloat Bytecode optimizer 66 1246 15 Antlr Parser generator 48 246 4 Summary Real world apps 80 avg 1000 avg 13 avg
36. Benchmark Characteristics Benchmark Function Heap Size in MB Allocation in MB % Nursery survival Xalan XML parser 108 980 17 Pmd Code analysis 98 364 23 Lusearch Text search 68 4294 4 Bloat Bytecode optimizer 66 1246 15 Antlr Parser generator 48 246 4 Summary Real world apps 80 avg 1000 avg 13 avg
37. Nursery and observer sizing nursery mature DRAM observer PCM mature meta data large mutator mutator large runtime meta data 4 MB 8 MB
38. PCM parameters Parameter Relative to DRAM Read latency 4X Write latency 10X Read power 0.9X Write power 4X Standby power ~0X Refresh power 0X
39. PCM lifetime model (1) First line of defence Hardware line wear levelling 0 1 2 3 4 N … 0 1 2 3 4 N
40. PCM lifetime model (1) First line of defence Hardware line wear levelling N 0 1 2 3 4 … (2) OS policy On page failure, give a new page to the JVM (2) Analytical model to compute lifetime # bytes × endurance # years = bytes_per_sec × 2∧25 0 1 2 3 4 N
41. Two baseline heap configurations nursery mature meta data large mutator mutator runtime GC DRAM nursery mature meta data large mutator mutator runtime GC PCM DRAM-Only PCM-Only
42. Proposed heap configuration (1) nursery mature meta data large mutator mutator runtime DRAM PCM KG-N
43. Proposed heap configuration (2) nursery mature DRAM observer PCM mature meta data large mutator mutator large runtime meta data KG-W
44. Simulated main memory systems Physical memory Heap Organization 32 GB DRAM DRAM-Only 32 GB PCM PCM-Only 1 GB DRAM + 32 GB PCM KG-N & KG-W
45. PCM-Only has an average lifetime of up to 2 years, making it impractical ! 0 10 20 30 4 GB 8 GB 16 GB 32 GB PCM life<me in years Main memory capacity PCM-only KG-N KG-W
46. KG-N improves PCM lifetime by 5X compared to PCM-Only 0 10 20 30 4 GB 8 GB 16 GB 32 GB PCM life<me in years Main memory capacity PCM-only KG-N KG-W
47. KG-W improves PCM lifetime by 11X compared to PCM-Only 0 10 20 30 4 GB 8 GB 16 GB 32 GB PCM life<me in years Main memory capacity PCM-only KG-N KG-W
48. PCM usage by KG-N & KG-W Avg Max KGN 92% 94% KGW 68% 88%
49. Comparison to prior work Page vs. object granularity
50. OS write partitioning (WP) moves frequently written pages to DRAM Rank 0 Threshold 1 2 3 1 2 4 8 P0 P0 P0 P0 2Rank
51. OS write partitioning (WP) moves frequently written pages to DRAM 0 1 2 3 1 2 4 8 P7 P6 P3 P0 P1 P2 P4 P5 Rank Threshold 2Rank PCM DRAM time promotion quantum
52. OS write partitioning (WP) moves frequently written pages to DRAM 0 1 2 3 1 2 4 8 P7 P6 P3 P0 P1 P2 P4 P5 Rank Threshold 2Rank time promotion quantum demotion quantum PCM DRAM
53. OS write partitioning (WP) moves frequently written pages to DRAM 0 1 2 3 1 2 4 8 P6 P3 P0 P1 P2 P4 P5 Rank Threshold 2Rank time promotion quantum demotion quantum PCM DRAM
54. Summary of OS write partitioning ² Keep often written pages in DRAM using a promotion mechanism ² Maximize the use of PCM using a demotion mechanism
55. WP results in 22% additional writes to PCM compared to KG-W 0 0.1 0.2 0.3 0.4 KG-N KG-W WP Writes normalized to PCM-Only 8 queues, 10 ms promotion, 50 ms demotion User + JVM Demotion
56. KG-N and KG-W reduces EDP by 36% over DRAM-Only 0.0 0.5 1.0 1.5 2.0 EDP rela<ve to DRAM-only DRAM-Only PCM-Only KG-N KG-W
57. KG-W improves performance by 30% over PCM-Only 1.0 1.5 2.0 Performance rela<ve to DRAM-only PCM-Only KG-N KG-W 1.4 1.7
58. PCM access latency dominate other overheads in KG-W 0 10 20 30 40 50 % Overhead Extra pointer tracking Observer GC Monitoring writes PCM latency Other
59. Performance, energy and lifetime: KG-W summary DRAM-Only PCM-Only Performance -40% +30% Energy-Delay +32% +35% Lifetime N/A 11X
60. Conclusions ² Monitor fine-grained write behaviour of objects in hybrid memory ² Exploit managed runtime for organizing objects in hybrid memory ² Use Kingsguard collectors to improve lifetime of upcoming hybrid memories