Cloud Data Center/Storage Power Efficiency Solutions Junyao Zhang 1

Current Storage Energy Efficiency Solutions: Tradeoff Energy/Performance  Multi-speed Disk: DRPM  CPU has Dynamic Voltage and Frequency Scaling (DVFS): can we use this idea on storage systems? 2  VM Consolidation: SRCMap  Request Consolidation (Replication-based): EERAID, Diverted Access  Redirect requests to some replicas to spin down the others  Data Consolidation: MAID, PDC  Skew data into partial disks/cache disks so that others can be shut down In what degree should we tradeoff energy/performance?

Power Proportional  A standard metric proposed by Google [1]:  Computer components should consume energy in proportion to the system utilization.  Observation: 3

Energy-proportional  First introduced by Google –”The case of energy proportional computing”  Computer components should consumes energy in proportion to the workload.  Observations:  Only 10%-50% utilization of servers and machines.  CPU’s consumption is slowing down, while storage and memory fraction is growing dramatically. (“significant improvements in memory and disk subsystems is needed”) 4

Storage power proportional(SPP)  “Energy Proportionality for Storage: Impact and Feasibility”  Exploited the benefit of PP in storage  Outlined techniques that can be involved in storage PP: consolidation, tiering/migration, write off-loading, adaptive seek speeds, workload shaping, opportunistic spindown, spindown/MAID, dedup/compression 5

Robust and Flexible Power- Proportional Storage 6

7

8

Solution 9  Fine-grained power proportionality for one data-set

10 Total number of blocks is B, [p+1,s] range corresponding to put the rest (r-1) replicas of the dataset

More 11

Read Performance 12

Write Performance 13

Handling Recovery  Bounded wake-up  Rebuild is power- proportional 14

Near power-proportional(cnt.) 15

Multi-data set: Fair Scheduling 16

Degradation 17

Sierra: Practical Power-proportionalilty for Data Center Storage  Power proportional layout with the concern of the following factors:  Fault-tolerance, Loading balance, Consistency, Good performance.  Three challenges:  Layout that allows significant power savings  Maintain read and write availability at the original levels  Predict the number of servers required at anytime 18

C1: Power-aware Layout 19  Gear 1 (g=1): need 2 nodes (Gear group 0) to keep 1 the copy of all nodes  Gear 2 (g=2): 4 nodes

C1: Power-aware Layout  Extending to three replicas and more: two options  Rack-aligned  Rotated 20

C2: Distributed virtual log (DVL)  Aim: maintain read/write availability  Write: if secondaries not available, entering “logging mode”(write primary replicas to DVL and replicate DVL r-1 times ) 21

C3: Gear Scheduler  Aim: predict system load and schedules servers to power down or up accordingly.  Observation: predict hourly behavior based on historical records of this hour. 22

Power Savings 23

Performance 24  Response Time

Conclusion  Power proportional is becoming an important metric for power/energy tradeoff  Rabbit proposed a idea-power proportional layout  Sierra considered factors such as: power, reliability, load balancing, consistency and etc. 25

SRCMap: Energy proportional storage using dynamic consolidation( 2010 FAST)  Basic idea: using a storage virtualization layer to enable energy proportionality by consolidating workloads to subset of physical groups.  Four interesting observations of storages:  The active data set for storage volumes is typically a small fraction of total used storage.  There is a significant variability in I/O workload intensity on storage volumes.  Data usage is highly skewed with more than 99% of the working set consisting of some ’really popular’ data and ’recently accessed’ data.  The read-idle time distribution of I/O workloads is dominated by small durations, typically less than five minutes. 26

SRCMap: Energy proportional storage using dynamic consolidation( 2010 FAST)  Server consolidation using Virtualization can achieve energy efficiency, but it needs to migrate data from spine down disk to active disk.  SRCMap is designed to solve this problem.  Design issues:  Multiple replica targets: flexibility to increase or decrease physical devices  1 mdisk for primary copy of vdisk  Sampling: creating full replica of vdisks (virtual disk)  only keep the working set of each vdisk.  Replica placement: not all replicas are created equal  place replicas unequal  Dynamic mapping: can not predetermine which volumes need to be active  active replica or secondary replica 27

SRCMap: Energy proportional storage using dynamic consolidation( 2010 FAST)  Design: 28

PARAID: A Gear-shifting Power-Aware RAID  Basic idea: use skewed striping pattern to adapt to the system load by varying active disks 29

Summarize  Similarities:  Though using different methods, all aiming to achieve storage energy proportionality  All consider reliability, fault tolerance, I/O scheduler and etc.  None of them considered data-affinity (Preliminary idea) 30

Preliminary idea(from pengju) 31

[1] L. A. Barroso and U. H¨olzle. The case for energy- proportional computing. Computer, 40(12):33–37, [2] E. Thereska, A. Donnelly, and D. Narayanan. Sierra: a powerproportional, distributed storage system. MSR-TR , November [3] H. Amur, J. Cipar, V. Gupta, G. R. Ganger, M. A. Kozuch, and K. Schwan. Robust and flexible power- proportional storage. In SoCC,

Thank you 33