Feng Chen and Xiaodong Zhang Dept. of Computer Science and Engineering The Ohio State University

Transcription

1 Caching for Bursts (C-Burst): Let Hard Disks Sleep Well and Work Energetically Feng Chen and Xiaodong Zhang Dept. of Computer Science and Engineering The Ohio State University

2 Power Management in Hard Disk Power management is a requirement in computer system design Maintenance cost Reliability & Durability Environmental effects Battery life Hard disk drive is a big energy consumer, for I/O intensive jobs e.g. hard disk drives account for 86% of total energy consumption in EMC Symmetrix 3000 storage systems. As multi-core CPUs become more energy efficient, disks are less. 2

3 Standard Power Management Dynamic Power Management (DPM) When disk is idle, spin it down to save energy When a request arrives, spin it up to service the request Frequently spin up/down a disk incurs substantial penalty: high latency and energy Disk energy consumption is highly hl dependent d on the pattern of disk accesses (periodically sleep and work) 3

4 Ideal Access Patterns for Power Saving Ideal Disk Power Saving Condition Requests to disk form a periodic and burst pattern Hard disk can sleep well and work energetically Disk accesses Long Disk Sleep Interval Increasing Burstiness of disk accesses is the key to saving disk energy time Disk Accesses in bursts 4

5 Buffer Caches Affect Disk Access Patterns Buffer caches in DRAM are part of hard disk service Disk data are cached in main memory y( (buffer cache) Hits in buffer caches are fast, and avoid disk accesses The buffer cache is able to filter and change disk access streams Applications requests Buffer Cache Prefetching Caching disk accesses Hard Disk 5

6 Existing Solution for Burst in Disks Forming Burst Disk Accesses with Prefetching Predict the data that are likely to be accessed in the future Preload the to-be-used data into memory Directly condense disk accesses into a sequence of I/O bursts Both energy efficiency and performance could be improved Limitations Buffer caching share the same buffer space. Energy-unaware replacement can easily change burst patterns created by prefetching Aggressive prefetching shrinks available caching space and demands highly effective caching Energy-aware caching policy can effectively complement prefetching No work has been done. Reference e Papathanasiou a as and Scott, USENIX 04 6

7 Caching can Easily Affect Access Patterns An example Original i Disk Accesses time Buffer Cache Solely relying on prefetching is sub-optimal, Bursty Disk Accesses organized by Prefetching Energy-aware caching policyis needed to create burst disk accesses Disk still cannot sleep well time a b c a b c 7 3 blocks to be evicted by Unfortunately, these Energy-Unaware blocks Caching to be accessed in a non-bursty way

8 Caching Policy is Designed for Locality Standard Caching Policies Identify the data that are unlikely to be accessed in the future (LRU) Evict the not-to-be-used data out from memory They are performance-oriented and energy-unaware Most are locality-based algorithms, e.g. LRU (clock), LIRS (clock-pro), Designed for reducing the number of disk accesses No consideration of creating burst disk access pattern C-Burst (Caching for Bursts) Our objectives: effective buffer caching To create burst disk accesses for disk energy saving To retain the performance (high hit ratios in buffer cache) 8

9 Outlines Motivation Scheme Design History-based C-Burst (HC-Burst) Prediction-based C-Burst (PC-Burst) Memory Regions Performance Loss Control Performance Evaluation vauato Programming Multimedia Multi-role l Server Conclusion 9

10 Restructuring Buffer Caches Buffer cache is segmented into two regions Priority region (PR) Hot blocks are managed using LRU-based scheme Blocks w/ strong locality are protected in PR Overwhelming memory misses can be avoided Retain the performance Energy-aware region (EAR) Cached blocks are managed using C-burst schemes Non-burst accessed blocks are kept here Re-accessed block (strong locality) is promoted into PR Buffer Cache Priority Region (LRU) Buffer Cache (LRU) Region size is dynamically adjusted Both performance and energy saving are considered Energy Aware Region (C-Burst) 10 Our focus in this talk

11 History-based C-Burst (HC-Burst) Distinguish different streams of disk accesses Multiple tasks run simultaneously in practice History record can help us to distinguish them Various tasks feature very different access patterns Burst e.g. grep, CVS, etc. Non-Burst e.g. make, mplayer, etc. Accesses reaching the hard disk is a mixture of both burst and non-burst accesses In aggregate, the disk access pattern is determined by the most non-burst one 11

12 Basic Idea of History-based C-Burst (HC-Burst) grep make bursty access long disk Idle intervals time A B C D E F G H I J K L non-bursty access short disk Idle Period To cache the blocks being accessed in a non-burst pattern, to reshape disk accesses to a burst pattern a b c d time e f g Buffer Cache Grep + make A B C D a 12 Aggregate Results: non-bursty access b c d E time F G H f g I J K L e

13 History-based C-Burst (HC-Burst) Identifying an application s access pattern I/O Context (IOC) An application s access history must be tracked across many runs Some tasks life times are very short Each task is associate with an IOC to track its data access history IOCs are maintained in a hash table for quick access Each IOC is identified by an I/O context ID (IOC ID) a hash value of executable s path or kernel thread s name 13

14 History-based C-Burst (HC-Burst) Epoch Application s access pattern may change over time Execution is broken into epochs, T seconds for each Too small or too large are both undesired Our choice T = (Disk Time-out Threshold ) / 2 No disk spin-down happens during one epoch with disk accesses Distribution of disk accesses during one epoch can be ignored grep epoch time A B C D E F G H I J K L make 14 a b c d time e f g

15 History-based C-Burst (HC-Burst) Block Group Blocks accessed during one epoch by the same IOC are grouped into a block group Each block group is identified by an process ID and an epoch time The size of a block group indicates the burtiness of data access pattern of one application The larger a block group is, the more bursty disk accesses are grep epoch Block Groups time A B C D E F G H I J K L make 15 a b c d time e f g

16 History-based C-Burst (HC-Burst) HC-Burst Replacement Policy Two types of blocks should be evicted Data blocks that are unlikely to be re-accessed Blocks with weak locality (e.g. LRU blocks) Data blocks that can be re-accessed with little l energy Blocks being accessed in bursty pattern Victim block group the largest block group Blocks that are frequently accessed would be promoted into PR Large block group often holds infrequently accessed blocks Blocks that are accessed in a bursty pattern stay in a large BG Large block group holds blocks being accessed in bursts 16

17 History-based C-Burst (HC-Burst) Multi-level Queues of Block Groups 32-level queues of block groups A block group of N blocks stays in queue Block groups on one queue are linked in the order of their epoch times Block groups may move upwards/downwards, if # of blocks changes The victim block group is always the LRU block group on the top queue w/ valid block groups Most Bursty (MB) Level 10 Victim block group LRU + MB Grep # of blks= Epoch ID = 10 make Make # of blks= Epoch # 8 Block is promoted to PR Block is demoted to EAR Burtiness Level 9 Level 1 Level 0 Least Bursty (LRU) Least Recent Used (LRU) 17 Epoch Time Most Recent Used (MRU)

18 Prediction-based C-Burst (PC-Burst) Main idea Certain disk access events are known and can be predicted. Evicting a block that is to be accessed during a short interval and close to a deterministic disk access With deterministic disk accesses and block reaccess time, selectively evicting blocks to be accessed in a short intervals and holding blocks to be accessed in long idle intervals short disk idle interval long disk idle interval deterministic disk accesses di d time predicted block accesses Block A Block B Holding Block B can avoid Breaking a long idle interval 18

19 Prediction-based C-Burst (PC-Burst) Prediction of Deterministic Disk Accesses Many well predictable periodic disk accesses exist in systems Timer-controlled OS events (e.g. pdflush) Multi-media workloads with steady consumption rate In practice, constant disk accesses may fluctuate sometimes System dynamics may affect disk I/O occasionally Real I/O pattern change may happen over time Challenge how to accommodate occasional system dynamics while responding quickly to real access pattern shift 19

20 Prediction-based C-Burst (PC-Burst) Prediction of Deterministic Disk Accesses Track each task s access history Each task is offered credits of [ -32, 32 ] 32 Feedback-based Prediction Compare observed interval with predicted interval Wrong prediction, reduce a task s credits Correct prediction Correct prediction, increase a task s credits Task w/ credit less than 0 is unpredictable Correct prediction Repeated mis-prediction increases the charge of credits exponentially Occasional dynamics only charge a task s credit slightly Real patter change quickly decreases a task s credits Wrong prediction credits

21 Prediction-based C-Burst (PC-Burst) Prediction of Block Re-Access Time Efficient Data Structure Block table maintains i 4 epoch times for all blocks, including non-resident blocks Conservative Prediction Only blocks having constant access intervals are believed predictable Logic Block Number (LBN)

22 Prediction-based C-Burst (PC-Burst) Multi-level Queues of Block Groups in PC-Burst 32-level queues of block groups each level has two queues Prediction Block Group p( (PBG) blocks to be accessed in the same future epoch time History Block Group (HBG) blocks being accessed in the same history epoch time Reference Points (RP) - resent deterministic disk accesses Victim block group The PBG on the top level,, in the shortest interval,, closest to a RP,, to be accessed in the furthest future If no PBG is found, search the same level queue of HBG Victim block group long Interval Shortest Interval Level 10 Level 0 22

23 Performance Loss Control Why there is performance loss? Increase of memory misses due to energy-oriented caching policy How to control performance loss Basic Rule Control the size of Energy Aware Region Perof. Perf. Loss Loss > < Tolerable Tolerable Rate Rate Shrink Enlarge Region Region Size size Main Memory Estimating gperformance loss Ghost buffer (GB) LRU replacement policy The increase of memory misses (M) Blocks not found in EAR, but found in GB The average memory miss penalty (P) Observed average I/O latency Performance loss L = M x P Priority Priority Region Region A memory miss Check Ghost Buffer Automatic tune Energy Aware Region size L < Tolerable Performance Loss Enlarge EAR region size L > Tolerable Performance Loss Shrink EAR region size Energy Ghost Aware Buffer Energy Region (LRU) Aware Region Hit in Ghost Buffer Perf. Loss ++ 23

24 Performance Evaluation Implementation Linux Kernel ,500 lines of code in buffer cache management and generic block layer Experimental Setup Intel Pentium 4 3.0GHz 1024 MB memory Western Digital WD160GB 7200RPM hard disk drive RedHat Linux WS4 Linux kernel Ext3 file system 24

25 Performance Evaluation Methodology Workloads run on experiment machine Disk activities are collected in Kernel on experiment machine Disk events are sent via netconsole to an monitor machine Disk energy consumption is calculated based on collected log of disk events using disk power models off line disk activities log Experiment machine Gigabit LAN Monitoring machine 25

26 Performance Evaluation Emulated Disk Models Hitachi DK23DA Laptop Disk IBM UltraStar 36Z15 SCSI Disk Hitachi DK23DA IBM UltraStar 36Z15 Capacity 30GB 18.4GB Cache 2MB 4MB RPM B/W 35MB/sec 53MB/sec Active Power 2 watt 13.5 watt Idle Power 1.6 watt 10.2 watt Standby Power 0.15 watt 2.5 watt Spin up 1.6 sec / 5 J 10.9 sec / 135 J Spin down 2.3 sec / 2.94 J 1.5 sec / 13 J 26

27 Performance Evaluation 27 Eight applications 3 applications w/ bursty data accesses 5 applications w/ non-bursty data accesses Three Case Studies Programming Multi-media Processing Multi-role servers Name Description MB/ epoch Request/epoch Make Linux kernel builder Vim Text editor Mpg123 Mp3 player Transcode Video converter TPC-H Database query # Grep Textual search tool Scp Remote copy tool CVS Version control tool

28 Performance Evaluation Case Study I programming Applications: grep, make, and vim Grep bursty workload Make, vim non-bursty workload C-Burst schemes protects data set of make from being evicted by grep Disk idle intervals are effectively extended 28

29 Performance Evaluation Case Study I programming Applications: grep, make, and vim Grep bursty workload Make, vim non-bursty workload C-Burst schemes protects data set of make from being evicted by grep Disk idle intervals are effectively extended 29 over 30% energy saving Nearly 0% intervals > 16 sec Over 50% intervals > 16 sec

30 Performance Evaluation Case Study II Multi-media Processing Applications: transcode, mpg123, and scp Mpg123 its disk accesses serve as deterministic accesses Scp bursty disk accesses Transcode non-bursty accesses PC-Burst achieves better performance than HC-Burst PC-Burst can accommodate deeper prefetching by efficiently using caching space Around 30% energy saving 30 Over 70% intervals > 16 sec

31 Performance Evaluation Case Study III Multi-role Server Applications: TPC-H # 17, and CVS TPC-H non-bursty disk accesses ( very random I/O ) CVS bursty disk accesses Dataset of TPC-H is protected in memory Performance of TPC-H is significantly improved No improvement on disk idle interval Reduced I/O latency over 35% energy saving 31

32 Our Contributions Design a set of comprehensive energy-aware caching policies, called C-Burst, which leverages the filtering effect of buffer cache to manipulate disk accesses Our scheme does not rely on complicated disk power models and requires no disk specification data, which means high compatibility to different hardware Our scheme does not assume using any specific disk hardware, such as multi-speed disks, which means our scheme is beneficial to existing hardware Our scheme provides flexible performance guarantees to avoid unacceptable performance degradation Our scheme is fully implemented in Linux kernel , and experiments under realistic scenarios show up to 35% energy saving with minimal performance loss 32

33 Conclusion Energy efficiency is a critical issue for computer system design Increasing disk access burstiness is the key to achieving disk energy conservation Leveraging filtering effect of buffer cache can effectively shape the disk accesses to an expected pattern HC-Burst scheme can distinguish different access pattern of tasks and create a bursty stream of disk accesses PC-burst scheme can further predict blocks re-access time and manipulate the timing of future disk accesses Our implementation of C-Burst schemes in Linux kernel and experiments show that C-Burst schemes can achieve up to 35% energy saving with minimal performance loss 33

34 34

35 References [USENIX04] A. E. Papathanasiou and M. L. Scott, Energy Efficient prefetching and caching. In Proc. of USENIX 04 [EMC99] EMC Symmetrix 3000 and 5000 enterprise storage systems product description guide

36 Memory Regions Buffer cache is segmented into two regions Priority region (PR) Hot blocks are managed using LRU-based scheme Blocks w/ strong locality are protected Overwhelming memory misses can be avoided Buffer Cache Priority Region 36 Energy-aware region (EAR) Cold blocks are managed using C-burst schemes Victim blocks are always reclaimed from EAR Re-accessed block is promoted into PR Accordingly, the coldest block in PR is demoted Promote a cold block Region size is tuned on line Both performance and energy saving are considered Insert a new block Energy Aware Region Demote a hot block Evict a cold block

37 Motivation Limitations Energy-unaware caching policy can significantly affect periodic bursty patternscreated by prefetching Improperly evicting a block may easily break a long disk idle interval Aggressive prefetching shrinks available caching space and demands highly effective caching Effective prefetching needs large volume of memory space, which raises high memory contention for caching Energy-aware caching policy can effectively complement prefetching When prefetching works unsatisfactorily, caching can give a hand by carefully selecting blocks for eviction 37

38 Motivation Prefetching Predict the data that are likely to be accessed in the future Preload the to-be-used data into memory Directly condense disk accesses into a sequence of I/O bursts Both energy efficiency and performance could be improved Caching Identify the data that are unlikely to be accessed in the future Evict the not-to-be-used data out from memory Traditional caching policies are performance-oriented rmance riented Designed for reducing the number of disk accesses No consideration of creating bursty disk access pattern 38

39 Prediction-based C-Burst (PC-Burst) Prediction of Deterministic Disk Accesses Track each task s access history and offer each task credits of [-32, 32] Feed-back based Prediction Compare observed interval with predicted interval If prediction is proved wrong, reduce a task s credits If prediction is proved right, increase a task s credits Task w/ credit less than 0 is unpredictable Repeated mis-prediction increases the charge of credits exponentially Occasional system dynamics only charge a task s credit slightly Real patter change quickly decreases a task s credits 39