1 Conserving Energy in RAID Systems with Conventional Disks Dong Li, Jun Wang Dept. of Computer Science & Engineering University of Nebraska-Lincoln Peter Varman Dept. of Electrical and Computer Engineering Rice University

2 References [1] S. Gurumurthi, A. Sivasubramaniam, M. Kandemir, and H. Franke, “DRPM: dynamic speed control for power management in server class disks,” ISCA’03 [2] D. Colarelli and D. Grunwald, “Massive arrays of idle disks for storage archives,” in Proceedings of Super Computing’ 2002 [3] E. Pinheiro and R. Bianchini, “Energy conservation techniques for disk array-based servers,” in Proceedings of the 18th International Conference on Supercomputing, 2004 [4] E. Varki, A. Merchant, J. Z. Xu, and X. Z. Qiu, “Issues and challenges in the performance analysis of real disk arrays,” IEEE Transactions on Parallel and Distributed Systems, [5] D. Li and J. Wang, “EERAID: Energy-efficient redundant and inexpensive disk array,” in Proceedings of 11 th ACM SIGOPS European Workshop, [6] D. Li, H. Cai, X. Yao, and J. Wang, “Exploiting redundancy to construct energy-efficient, high-performance RAIDs,” Tech. Rep. TR , Computer Science and Engineering Department, University of Nebraska Lincoln, 2005.

3 Outline Introduction Motivation eRAID Design Evaluation Leveraging eRAID Conclusions

4 Introduction Energy-efficient storage system, total cost of ownership (TCO), … Short request inter-arrival time Long disk state switch time of conventional disks Current solutions: multi-speed disks[1] Create long idle period for conventional disks  unbalance workloads Two approaches  Relocating data: MAID[2], PDC[3]  Redirecting requests: EERAID[5] Introduction Motivation Evaluation Conclusions Leveraging Design

5 Motivation Major limitations of state of the art  few workable solutions for conventional disk based systems  single performance measurement  no differentiation of workload time criticality Three observations  redundant information of RAID systems  spare service capacity  queueing model Introduction Motivation Evaluation Conclusions Leveraging Design

6 eRAID Design Main idea  spin down, partially or entirely, mirror disks to standby  read, write Features  soft solution --- no hardware change  consider two performance metrics Research issue  maximize energy saving  without violating predefined performance degradation limits for both throughput and response time  assume workloads have little change between two consecutive time windows Introduction Motivation Evaluation Conclusions Leveraging Design

7 Solving for Performance Degradation Our approach: using queueing models to do predictions 1. model RAID-1 system and get performance measures 2. examine how the input parameters are changed 3. get new performance measures with changed input parameters 4. compare these two results Four workloads: synchronous read (SR), asynchronous read (AR), synchronous write (SW) and asynchronous write (AW) Real system: HP SureStore E Disk Array FC60 Introduction Motivation Evaluation Conclusions Leveraging Design

8 Read Load Models Introduction Motivation Evaluation Conclusions Leveraging Design

9 Read Load Performance Computing The possible changes of input parameters:  disk access probability  disk service time --- negligible Synchronous read load:  Mean Value Analysis (MVA) technique  eRAID --- double access probabilities of corresponding primary disks Asynchronous read load:  no throughput degradation for stable systems  eRAID --- double work loads of corresponding primary disks Introduction Motivation Evaluation Conclusions Leveraging Design

10 Write Load Model Introduction Motivation Evaluation Conclusions Leveraging Design Controller cache  write back policy FC60: two-threshold write back policy  destage_threshold, max_ditry Disk array: M/M/1/K queueing model[4]

11 Write Load Performance Computing Dirty data arrival rate d  SW load: d =  * cache_miss_rate  : max throughput with infinite cache size  AW load: d = * cache_miss_rate independent with the system The possible changes of input parameters:  service rate: N/2 => (N-2i)/2  maximum queue length  cache miss rate --- unnoticeable Introduction Motivation Evaluation Conclusions Leveraging Design

12 Solving for Energy Saving N-disk RAID1 Time window length T Request number R Mean service time t Asyn. load:  2 =  1 Sync. load:  2 <  1 Introduction Motivation Evaluation Conclusions Leveraging Design E eRAID = E active +E idle +E standby +E switch (N-i) disks i disks N disks E base = E active +E idle

13 Control Algorithm Time-window Solve multi-constraint problem:  select LFU disks Conservative control Introduction Motivation Evaluation Conclusions Leveraging Design

14 Evaluation Disk power model: IBM Ultrastar 36Z15 Simulator: augmented Disksim Traces: Cello99 and TPC-C20 8-disk RAID1 system Two scenarios Introduction Motivation Evaluation Conclusions Leveraging Design

15 Preliminary Results Introduction Motivation Evaluation Conclusions Leveraging Design CASE I: Limit T  10% & Limit X  3% LoadOverall D T Overall D X Overall S E AR7.5%0.0%10.2% SR7.1%1.5%11.8% AW4.3%0.0%13.3% SW0.0% CASE II: Limit T  50% & Limit X  6% LoadOverall D T Overall D X Overall S E AR29.5%0.0%30.0% SR25.9%4.6%27.7% AW14.3%0.0%23.5% SW41.4%1.6%7.4%

16 Leveraging eRAID Associate a load threshold f (1/2<f<1) for each disk  when primary disk load exceeds f, spin up mirror disk to share the load  conventional mirrored layout: spin up one mirror disk  our new layout: spin up less than one mirror disk Layout files of one primary disk to a set of mirror disks Introduction Motivation Evaluation Conclusions Leveraging Design

17 An example: N=10 and f=2/3 Introduction Motivation Evaluation Conclusions Leveraging Design

18 Conclusions An energy saving policy, eRAID, for conventional disk based RAID-1 systems 30% energy-saving without violating predefined performance constraints A new data layout scheme for further energy-saving Limitations  circumscribed by the accuracy of queueing models  approximated input parameters, e.g. process number and mean process delay  conservative control Introduction Motivation Evaluation Conclusions Leveraging Design

19 Thank you! Questions?

20 Creating Disk Idle Period in RAID-5: An Example 4-disk RAID 5 system A parity group containing data stripe 1, 2, 3 and parity stripe p that are saved in disk 1, 2, 3 and 4 respectively There is a read request for stripe 1. To service such a read, we could either read stripe 1 from disk 1, or read stripe 2, 3 and p, then calculate stripe 1 on the fly by an XOR calculation. More details can be found in our technical report[6]