Evaluating Phase Change Memory for Enterprise Storage Systems

Transcription

1 Hyojun Kim Evaluating Phase Change Memory for Enterprise Storage Systems IBM Almaden Research

2 Micron provided a prototype SSD built with 45 nm 1 Gbit Phase Change Memory Measurement study Performance Characteris?cs Part I Storage Simula?on Part II

3 3 Part I. PCM performance

4 Flash Memory vs. Phase Change Memory Flash memory traps charges to remember bits PCM melts and cools material to remember bits Temperature Reset (Amorphous phase) Set (Crystalline phase) Time 4

5 PCM write latency Temperature Reset (amorphous) Set (Crystalline) 150 ns ISCA 09, SOSP 09, FAST 11: 150 ns was used as PCM write latency FAST Projec=ons: 5

6 PCM write throughput and power consumption PCM Bit Cell PCM PCM PCM PCM PCM PCM chips have limited write throughput because of limited power budget 6

7 Example: VLSI Circuits, 2004, F. Bedeschi 7

8 8

9 Write latency and throughput comparison! Write latency PCM SET opera?on?me: 200 ns Flash page program?me: 200 μs! Write throughput per chip PCM: 4 bits / 200 ns = 2.5 MB/s Flash: 4 KB / 200 us = 20.5 MB/s 9

10 PCM technology survey paper (2013) 10

11 Clarifications about PCM performance! Write latency Write throughput! 150 ns SET?me = Material performance! Material performance Chip performance! Chip performance SSD performance 11

12 All PCM SSD vs. emlc SSD! 45 nm PCM! 64 GB capacity! emlc NAND flash! 1.8 TB capacity 12

13 Fine-grained I/O latency measurement Workload Generator Linux Storage So`ware stack Sta?s?cs Collector Measure Fine- grained I/O latency Device Driver 13

14 4KB Random Read Latency (5M samples) Percentage PCM Mean 6.7µs Onyx: 38 μs Latency (µs) Maximum 194.9µs Standard deviation 1.5µs 6.7 μs x shorter average latency μs Percentage Maximum 54.7ms Standard deviation 76.2µs emlc Latency (µs) Mean 108.0µs

15 4KB Random Write Latency (1M samples) 2.5 Percentage PCM Onyx: 179 μs Maximum 378.2µs Standard deviation 2.2µs Mean 128.3µs Latency (µs) 37.1 μs 3.4x longer μs Percentage emlc Maximum 17.2ms Standard deviation 153.2µs Mean 37.1µs Latency (µs)

16 Summary of Performance Numbers PCM SSD emlc SSD 4KB Read Latency 6.7 μs μs 4KB Write Latency μs 37.1 μs! For read, PCM SSD is about 16x faster! For write, PCM SSD is about 3.4x slower than emlc SSD using DRAM as write buffer 16

17 17 Part II. Storage simulation Will PCM be useful for enterprise storage systems?

18 PCM / Flash / HDD multi-tiered storage PCM SSD emlc SSD 15K RPM HDD 4KB Read Latency 6.7 μs μs 5 ms 4KB Write Latency μs 37.1 μs 5 ms Normalized Cost Hot data Flash Hot / Read intensive Hot / Write intensive PCM Cold data HDD Cold data 18

19 Multi-tiered storage simulation 1. Assume mul?-?ered storage made of X % of PCM, Y % of Flash, Z % of HDD 2. Es?mate Performance (IOPS) 3. Es?mate Storage Cost (Normalized: HDD = 1) 4. Evalua?on metric: Es?mated Performance (IOPS) Es?mated Cost ($) 19

20 Tiering simulation methods! Sta?c- op?mal data placement Complete knowledge about I/O workload No data movement! Dynamic?ering Reac?ve data movement based on I/O traffic 20

21 Retail Store, 2012 June, one week duration Cumulative amount (%) Read TiB Write 45.0 TiB Amount of Read Amount of Write Portion (%) of total accessd capacity (16.1 TiB) 21

22 Retail Store: IOPS/$ with Dynamic tiering 3000 Best: PCM(22%),Flash(78%) IOPS/$ PCM Percentage Flash Percentage

23 Retail Store 3,500 " 3,220 " IOPS/$" 3,000 " 2,500 " 2,000 " 1,500 " 1,000 " 500 " 200 " 1,713 " 1,661 " Static optimal data placement Dynamic tiering 2,757 " 61% - " HDD 100%" Flash 100%" PCM 100%" PCM 30% Flash 67% HDD 3%" PCM 22% Flash 78%" 23

24 Bank Telecommunication

25 Applica?on Server IBM Easy Tier Server EMC XtremCache NetApp FlashAccel FusionIO IO Turbine Flash Cache Shared Storage 25

26 Application server side cache simulation! IO by IO storage traces From real customer s produc?on systems (over 24 hour dura?on) Manufacturing, media, medical companies! Cache Simula?on methods Write- through, LRU replacement Evalua?on metric: average read latency! Performance parameters 26

27 Same cost cache configurations PCM (16G) Flash (64G) 4G 8 G Flash (32G) Flash (48G) Flash 64G PCM 16G P 8G + F 32G P 4G + F 48G Flash 128G PCM 32G P 16G + F 64G P 8G + F 96G Flash 256G PCM 64G P 32G + F 128G P 16G + F 192G 27

28 Manufacturing: I/O distribution (24 hours) Cumulative amount (%) Write 3.8 TiB Read 1.1 TiB Amount of Read Amount of Write Portion (%) of total accessd capacity (246.5 GiB) 28

29 Manufacturing: cache simulation results Average read latency (μs)" 350.0" 300.0" 250.0" 200.0" 150.0" 100.0" 50.0" 187.7" 308.5" 241.9" 242.7" 151.0" 176.2" 153.3" 157.0" 141.4" 104.5" 87.2" 106.8" 0.0" 29

30 Media Medical 30

31 31 Summary and Conclusion

32 Summary! Let s be careful to pick right performance number for PCM! Performance measurement results PCM SSD emlc SSD 4KB Read Latency 6.7 μs μs 4KB Write Latency μs 37.1 μs! Storage simula?on results show About 12-66% improved IOPS/$ for?ered storage Up to 35% reduced average read latency for server caching 32

33 Concluding question Phase Change Memory for enterprise storage Silver bullet or Snake oil? What do you think? 33

34 Additional experiments! The PCM SSD was shipped to Almaden, finally! Real experiments with Sysbench OLTP benchmark Read- only workload Single thread test / 8 thread test 34

35 Sysbench OLTP benchmark, read only 1 thread: emlc SSD Transactions per second (TPS) Flash (156.8) Time (minutes) 35

36 Sysbench OLTP benchmark, read only 1 thread: emlc SSD vs. PCM SSD 250 PCM (219.0) Transactions per second (TPS) Flash (156.8) Time (minutes) 36

37 Sysbench OLTP benchmark, read only 8 thread: emlc SSD Flash (754.7) Transactions per second (TPS) Time (minutes) 37

38 Sysbench OLTP benchmark, read only 8 thread: emlc SSD vs. PCM SSD Flash (754.7) Transactions per second (TPS) PCM (753.3) Time (minutes) 38

39 PCM caching over emlc flash SSD 250 PCM Flash flashcache/30g SCE/30G 200 Transactions per second (TPS) Time 39

40 40 Thank you