Caching less for better performance: Balancing cache size and update cost of flash memory cache in hybrid storage systems"

Transcription

1 Caching less for better performance: Balancing cache size and update cost of flash memory cache in hybrid storage systems" Yongseok Oh" Jongmoo Choi! University of Seoul! Donghee Lee! Dankook University! Sam H. Noh! Hongik University! 1

2 Hybrid Storage Systems Harness benefits of SSDs and HDDs! High performance, large capacity, affordable cost! SSDs used as flash cache (NVCache)! Seagate Momentus XT(SLC 4GB), OCZ RevoDrive Hybrid (MLC 100GB)! Our focus: issue of managing flash cache! SSD Hybrid Storage System! HDD + High performance! Low power consumption Large Capacity! Low cost!! 2

3 Important Characteristics of Flash based SSDs Maintain Over-Provisioned Space (OPS)! Reserved space for Garbage Collection (GC)! Greatly influence GC performance! Typical SSDs! OPS size is fixed! Optimal size is unknown! Cannot adapt to workload changes! Fixed Size OPS! Caching Space" OPS" Flash based SSD! 3

4 Our Goal: Find Optimal OPS Size Overall cost! (e.g., response time of hybrid storage)! Performance! Point of Optimal Performance! (Our Goal)! Cache miss rate! GC cost! Less! Over-Provisioned Space More Caching Space" OPS" Caching Space" OPS" 4

5 Workload Dependent Optimal Partitioning Idea" Workload Dependent Optimal Partitioning! Hit! Perf.! Cachi ng Space" OPS" Caching Space Size! Flash Cache Caching Space" OPS" Flash Cache HDD" Proposed Hybrid Storage! Hit! Optimal OPS size changes! Caching Space Size! Perf.! Caching Space" Flash Cache OPS" Periodically adjust OPS size to maximize the performance! Based on hit ratio and garbage collection cost! Question: how to find optimal OPS size?! Solution: Hybrid Storage Cost Model (Dynamically adjusted according to workload)! 5

6 Outline Introduction! Hybrid Cost Model! Implementation! Evaluation! Conclusion! 6

7 OS 101: Access Cost Model (ACM) C ACM = Hit Rate x Cache Cost + (1-Hit Rate) x Miss Penalty! Expected I/O cost! Performance! Request! Hit! Buffer Cache HDD" Storage Hierarchy! Miss! Capacity! 7

8 Hybrid Storage: Access Cost Model C ACM (u)= Hit Rate(u) x Flash Cache Cost(u) + (1-Hit Rate(u)) x Miss Penalty(u)! Request! Hit! Miss! u! 1-u! Flash Caching Space OPS Cache! HDD" C ACM (u) represents expected I/O cost based on u! Incorporating u into the access cost model! Flash Cache is divided based on u (tunable)! u is fraction of caching space in flash cache (e.g., 0 u 1.0)! u influences hit ratio and access cost of flash cache! 8

9 Overview of Hybrid Cost Model Hybrid cost model represents expected I/O cost! Combines hybrid read cost model and hybrid write cost model! Caching space divided into read and write spaces! For this talk we derive hybrid read cost model! Hybrid Cost Model! Hybrid Read Cost Model! Hybrid Write Cost Model! Flash Cache! Read Write! OPS HDD" 9

10 Overview of Hybrid Cost Model Hybrid cost model represents expected I/O cost! Combines hybrid read cost model and hybrid write cost model! Caching space divided into read and write spaces! For this talk we derive hybrid read cost model! Hybrid Cost Model! Hybrid Read Cost Model! Flash Cache! Read Read Write! OPS HDD" 10

11 OPS Aware Hybrid Read Cost Model C HR (u) = Hit Rate(u) x Flash Cache Cost(u) + (1-Hit Rate(u)) x Miss Penalty(u)! Read Hit Rate! Flash Read! Hit! HDD Read + Flash Write! Miss! u! 1-u! Maintain read data from HDD! Flash Read Cache OPS Cache! Requirements for derivation! Read Hit Rate Function! HDD Cost Model! Flash Cache Cost Model! HDD" 11

12 Hybrid Read Cost Model C HR (u) = Hit Rate(u) x Flash Cache Cost(u) + (1-Hit Rate(u)) x Miss Penalty(u)! Read Hit Rate! Flash Cache Read! HDD Read + Flash Cache Write! Read Hit Rate Function! HDD Cost Model! Flash Cache Cost Model! Finding Optimal Point! 12

13 Read Hit Rate Function Read Hit rate function: H R (u), miss rate: 1-H R (u)! Related to workload pattern! Depends on u! Hit! Miss! Flash Cache! Read Cache u=0.1! OPS H R (0.1) Hit Rate! Low! u! Flash Cache! Hit! Miss! Read Cache OPS u=0.9! H R (0.9) Hit Rate u High! 13

14 Hybrid Read Cost Model C HR (u) = Hit Rate(u) x Flash Cache Cost(u) + (1-Hit Rate(u)) x Miss Penalty(u)! H R (u)! Flash Cache Read! HDD Read + Flash Cache Write! Read Hit Rate Function! HDD Cost Model! Flash Cache Cost Model! Finding Optimal Point! 14

15 HDD Cost Model HDD I/O requires positioning cost + bus transfer cost [Hylog]! HDD Read: C DR = C D_RPOS + P/B! HDD Write: C DW = C D_WPOS + P/B! Independent from u! Hit! Miss! u! 1-u! Read Cache OPS Read HDD" Notation C D_RPOS C D_WPOS P B Description Read positioning Cost Write positioning Cost Page size (in bytes) Bandwidth 15

16 Hybrid Read Cost Model C HR (u) = Hit Rate(u) x Flash Cache Cost(u) + (1-Hit Rate(u)) x Miss Penalty(u)! H R (u)! Flash Cache Read! C DR + Flash Cache Write! Read Hit Rate Function! HDD Cost Model! Flash Cache Cost Model! Read Cost Model! Write Cost Model! Finding Optimal Point! 16

17 Flash Cache Read Cost Model Hit request requires flash page read: C PR! Near constant cost (e.g., 25us)! Regardless of garbage collection cost! Independent from u! Hit! Flash Read! Miss! u! 1-u! Read Cache OPS HDD" 17

18 Hybrid Read Cost Model C HR (u) = Hit Rate(u) x Flash Cache Cost(u) + (1-Hit Rate(u)) x Miss Penalty(u)! H R (u)! C PR! C DR + Flash Cache Write! Read Hit Rate Function! HDD Cost Model! Flash Cache Cost Model! Read Cost Model! Write Cost Model! Finding Optimal Point! 18

19 Flash Cache Write Cost Model Miss request requires flash page write: C PW (u)! Write cost + GC cost(u)! GC cost(u) varies depending on u [LFS, Janus-FTL]! As u increases, GC cost(u) increases C PW (u) increases! Hit! Miss! Detailed Derivation! Flash Write! u! 1-u! Read Cache OPS C GC (u) = u N NP C CP + C E C PW (u) = C (u) GC + C PROG (1 u) N P See the paper for derivation! HDD" 19

20 Hybrid Read Cost Model C HR (u) = Hit Rate(u) x Flash Cache Cost(u) + (1-Hit Rate(u)) x Miss Penalty(u)! Derive! C HR (u) = H R (u) * C PR + (1-H R (u)) * (C DR +C PW (u))! Read Hit Rate Function! HDD Cost Model! Flash Cache Cost Model! Finding Optimal Point! 20

21 Finding Optimal Point 1. Observe Hit Ratio! 3. Find Optimal Point! Optimal Point" u=0.92" (a) Read hit ratio (b) Read access cost 4. Adjust! 2. Calculate for all values of u! C HR (u) = H R (u) * C PR + (1-H R (u)) * (C DR +C PW (u))! Flash Cache: e.g., 4GB Partition based on optimal u = 0.92! Caching Space 3.68GB OPS 0.32GB 21

22 Hybrid Cost Model: Distinguishing Read and Write C HY (u, r) represents expected I/O cost based on u and r! Caching space divided into read and write spaces based on r! r is fraction of read space in caching space (e.g., 0 r 1.0)! Modification: C HR (u) C HR (u, r), C HW (u) C HW (u, r)! Used to find optimal values: u and r! Hybrid Read Cost Model: C HR (u, r)! Hybrid Cost Model: C HY (u, r)! Hybrid Write Cost Model: C HW (u, r)! See the paper for derivation! Read r u Write 1-r 1-u OPS 22

23 Calculate Hybrid Cost Model 1. Observe Hit Ratio! 3. Draw Access Cost Graph! Better! (a) Read hit ratio (b) Write hit ratio (c)expected access cost 2. Calculate based on u and r! Hybrid Cost:! Hybrid Read Cost:! Hybrid Write Cost:! C HY ( u, r) = C ( u, r) IO + C ( u, r) IO HR C HR (u,r) = H R (u, r) C PR + (1 H R (u,r)) (C DR + C PW (u)) R C HW (u, r) = H W (u, r) C WH + (1 H W (u, r)) (C PR + C DW + C PW (u)) HW W 23

24 Optimal Partitioning Algorithm with Hybrid Cost Model Periodically Execute Optimal Partitioning Algorithm! for u step; u < 1.0; u u + step do for r 0.0; r 1.0; r r + step do cur cost C HY (u, r) if cur cost < op cost then op cost cur cost op u u, op r r end if end for end for ADJUST CACHE SIZE(op u, op r) end procedure Find u and r resulting in Optimal I/O Cost! Optimal Point at op_u=0.64, op_r=0.25! r! Read 0.64GB op_u! Write 1.92GB! op_r! Adjust Flash Cache partition! Flash Cache: e.g., 4GB op_r = 0.25! op_u = 0.64! 0.75! 0.36! OPS 1.44GB 24 u!

25 Outline Introduction! Hybrid Cost Model! Implementation! Evaluation! Conclusion! 25

26 Optimal Partitioning Flash Cache Layer (OP-FCL) I/O request arrives Seq. I/O Detector" If identify,! go to HDD! If non-seq. I/O,! go to Flash Cache! Cache Miss! Page Replacer Cache Hit! Read LRU! Write LRU! Mapping Manager" Translation Table Logical to Physical! Workload Tracker" Hit! Curves! Shrink! Enlarge! Periodically Execute! Partition Resizer Workload Dependent Optimal Partitioning! HDD Flash Cache 26

27 Adapt to Workload Pattern Read Hit! u! Write Hit! u! Read Write OPS Flash Cache! Workload changes! Read Hit! u! Write Hit! u! Resize! Invalidate! Destage! Read Hit! u! Workload changes! Write Hit! Enlarge! Resize! u! 27

28 Outline Introduction! Hybrid Cost Model! Implementation! Evaluation! Conclusion! 28

29 Evaluation Setup Hybrid Storage Simulator! CMU DiskSim 4.0 and MSR SSD extension! Flash Cache Layers (FCLs)! Fixed Partitioning (FP-FCL) - Fixed size OPS - Typical SSD product! Read Write (RW-FCL) - Fixed size OPS - Distinguishes read and write! Optimal Partitioning (OP-FCL) Dynamically adjusted!! based on workload! Configurations! Config. 1: 4GB flash cache + 10K RPM HDD! Config. 2: 16GB flash cache + three 10K RPM HDDs! 29

30 Workload Traces Financial [UMass] with Config. 1! Random write dominant! OLTP application running at a financial institutions! Search Engine [UMass] with Config. 1! Random read dominant! Web search engine! Exchange [SNIA] with Config. 2! Random read/write mixed! Microsoft employee server! Home [FIU] with Config. 1! Development, testing, and plotting in NFS Server! MSN [SNIA] with Config. 2! MSN storage back-end file store! 30

31 Response Time Results Mean Resp. Time (ms) FP-FCL RW-FCL OP-FCL Caching Space (%) in SSD Mean Resp. Time (ms) FP-FCL RW-FCL OP-FCL Caching Space (%) in SSD Mean Resp. Time (ms) FP-FCL RW-FCL OP-FCL Caching Space (%) in SSD (a) Financial! (b) Search Engine! (c) Exchange! OP-FCL shows near-optimal performance! Optimal performance depends on workload characteristics! 31

32 Response Time Results Mean Resp. Time (ms) FP-FCL RW-FCL OP-FCL 50%! 95%! 70%! Caching Space (%) in SSD Mean Resp. Time (ms) FP-FCL RW-FCL OP-FCL Caching Space (%) in SSD Mean Resp. Time (ms) FP-FCL RW-FCL OP-FCL Caching Space (%) in SSD (a) Financial! (b) Search Engine! (c) Exchange! OP-FCL shows near-optimal performance! Optimal performance depends on workload characteristics! 32

33 Dynamic Adjustment Cache Size (GB) OPS! Cache Size (GB) OPS! Write! 1 Read! Write! Read! 0 Logical Time Read! Logical Time Logical Time (a) Financial! (b) Search Engine! (c) Exchange! Cache Size (GB) OPS! OP-FCL dynamically adjusts cache spaces according to workloads! Financial and Exchange! Considerable OPS is used to lower garbage collection cost" Search Engine! Most caching space is used to maintain read data! 33

34 Dynamic Adjustment Cache Size (GB) OPS! Cache Size (GB) OPS! Write! 1 Read! Write! Read! 0 Logical Time Read! Logical Time Logical Time (a) Financial! (b) Search Engine! (c) Exchange! Cache Size (GB) OPS! OP-FCL dynamically adjusts cache spaces according to workloads! Financial and Exchange! Considerable OPS is used to lower garbage collection cost! Search Engine! Most caching space is used to maintain read data" 34

35 Effect on Lifetime of Flash Cache Lifetime Optimal! Performance Optimal! Average Erase Count FP-FCL RW-FCL OP-FCL Caching Space (%) in SSD Average Erase Count FP-FCL RW-FCL OP-FCL Caching Space (%) in SSD Average Erase Count FP-FCL RW-FCL OP-FCL Caching Space (%) in SSD (a) Financial! (b) Search Engine! (c) Exchange! Lifetime of flash cache is an important issue! Optimal point of lifetime differs from that of performance! Our focus is to improve the performance of flash cache! Optimizing lifetime of flash cache left as future work! 35

36 Conclusion Trade-off exists! Caching benefit vs update cost! We proposed OP-FCL for Hybrid Storage Systems! Use workload dependent cost model! Adjust read, write, and OPS sizes based on proposed cost model! Show near-optimal performance compared to others! Future direction! Develop better destaging and replacement algorithm! Make SSD lifetime aware hybrid storage system! 36

37 Caching less for better performance: Balancing cache size and update cost of flash memory cache in hybrid storage systems Yongseok Oh! Jongmoo Choi! University of Seoul! Donghee Lee! Dankook University! Sam H. Noh! Hongik University! 10th USENIX Conference on File and Storage Technologies (FAST 12)! 37