energy efficient prefetching – from models to implementation
DESCRIPTION
With the rapid growth of the production and storage of large scale data sets it is important to investigate methods to drive the cost of storage systems down. We are currently in the midst of an information explosion and large scale storage centers are increasingly used to help store generated data. There are several methods to bring the cost of large scale storage centers down and we investigate a technique that focuses on transitioning storage disks into lower power states. This talk introduces a model of disk systems that leverages disk access patterns to prefetch popular sets of data to produce energy saving opportunities. Using the model we have developed a simulator that allows us to quickly change various parameters to investigate the relationship that file access patterns, disk energy parameters, and simulation parameters have on the overall energy efficiency of disk systems. To help improve the validity of our simulation results we leveraged the validated disk simulator, DiskSim, and added disk power models to DiskSim. This allowed us to test our energy efficient strategies with a validated storage system simulator. The last part of this talk focuses on implementing a large scale storage system virtual file system. We introduce the Energy Efficient Virtual File System, or EEVFS, to mange the data placement and disk states in a cluster storage system. Our modeling and simulation results indicated that large data sizes and knowledge about the disk access pattern are valuable for storage system energy savings techniques. Storage servers that support applications that stream media is one key area that would benefit from our strategies. The final idea introduced in this talk is the concept of parallel striping groups, which attempt to improve the performance of EEVFS while maintaining energy savings.TRANSCRIPT
Energy Efficient Prefetching – from models to Implementation
04/11/23 1
Adam Manzanares and Xiao Qin
Department of Computer Science and Software Engineering
Auburn Universityhttp://www.eng.auburn.edu/~xqin
Adam Manzanares
Ph.D. May 2010.
About me
Ph.D.’04, U. of Nebraska-Lincoln
04-07, New Mexico Tech 07-10, Auburn University
About My Research Group
Presentation Outline
• Motivation
• Modeling Work
• DiskSim Modifications
• Energy Efficient Virtual File System (EEVFS)
• Parallel Striping Groups in EEVFS
• Conclusion
04/11/23 5
MotivationEPA Report to Congress on Server and Data Center Energy Efficiency, 2007
04/11/23 6
Motivation Using 2010 Historical Trends Scenario
◦ Server and Data Centers Consume 110 Billion kWh per year
◦ Assume average commercial end user is charged 9.46 kWh
◦ Disk systems can account for 27% of the energy cost of data centers
04/11/23 7
Buffer Disk Architecture
RAM BufferRAM Buffer
m buffer disksm buffer disks n data disksn data disks
Buffer Disk ControllerBuffer Disk Controller
Data Partitioning Data Partitioning
Security Model Security Model
Load BalancingLoad Balancing
Power ManagementPower Management
PrefetchingPrefetching
Disk RequestsDisk Requests
Energy-Related Reliability Model Energy-Related Reliability Model
04/11/23 8
IBM Ultrastar 36Z15
04/11/23 9
Transfer Rate 55 MB/s Spin Down Time: TD 1.5 s
Active Power: PA 13.5 W Spin Up Time: TU 10.9 s
Idle Power: PI 10.2 W Spin Down Energy: ED 13 J
Standby Power: PA 2.5 W Spin Up Energy: EU 135 J
Break-Even Time: TBE 15.2 S
Prefetching
Disk 1
Disk 2
Disk 3
Buffer Disk
04/11/23 10
Why Modeling & Simulation
• Allows us to determine the potential of our research ideas
• Can quickly evaluate many simulation parameters
• Allows us to test architectures and hardware without having the physical resources
04/11/23 11
Modeling & Simulation Work
Developed Mathematical Model◦Disk Energy Consumption◦Conditions to prefetch
Developed Energy Saving Principles◦Investigated cases that exploit the energy
saving principles Implemented model in JAVA based simulator
04/11/23 12
Energy Saving Principles
Energy Saving Principle One◦Increase the length and number of idle
periods larger than the disk break-even time TBE
Energy Saving Principle Two◦Reduce the number of power-state
transitions
04/11/23 13
Paramaters TestedParameter Values
Data Size 1, 5, 10, 25 MB
# of Data Disks 4, 8, 12
Inter-arrival Delay 0, 0.1, 0.5, 1 S
Hit Rate 85, 90, 95, 100%
04/11/23 14
Energy Savings Hit Rate 85%
04/11/23 15
State Transitions
04/11/23 16
Parameter Generalizations
• Larger data sizes produce greater energy savings and less state transitions
• Increasing the inter-arrival delay increases energy savings
• More data disks per buffer disks increases energy efficiency
• High hit rates produce the greatest energy efficiency
04/11/23 17
Modeling & Sim. Summary
Hit Rate, Inter-arrival Delay, & Data Size combine to produce Idle Windows
Transitions important to reduce energy consumption◦ May increase/decrease to reduce energy consumption
Disk parameters have large impact on energy savings
Model and simulator developed in-house
04/11/23 18
DiskSim
• Event driven simulator developed at CMU
• Simulates disks at the block level
• The simulator has been validated
• Discrete event based simulator
• Provides a large amount of statistics
• Lacks Disk Power Models
• Ability to simulate large storage systems
04/11/23 19
File System Simulator
• Large files important to energy savings
• Popularity of data is also useful
• Developed a block to file translator
• Interacts with DiskSim
04/11/23 20
DiskSim with File System Simulator
04/11/23 21
Modified DiskSim Results
04/11/23 22
Modified DiskSim Summary
• Provides us with accurate disk statistics
• Only the changes to DiskSim need to be validated
• Heavily dependent upon disk parameters
• May miss details that can only be found in implementation
04/11/23 23
Why a Cluster File System
• Block level prefetching difficult
• Natural place to track file accesses
• Control placement of data among storage nodes, and data disks
• Tiered approach simplifies management of files and disk states
• Eliminates some shortcomings of modeling and simulation
04/11/23 24
Energy Efficient Virtual File System
04/11/23 25
EEVFS Process Flow
04/11/23 26
EEVFS Testbed
Parameter Storage Server Storage Node Type 1
Storage Node Type 2
CPU P4 2.0 GHz P4 3.2 GHz P4 2.4 GHz
Memory (MB) 2000 1000 512
Network Interconnect
1000 1000 100
Disk Type SATA ATA/133 ATA/133
Disk Capacity 120 GB 80 GB 80 GB
Disk Bandwidth 100 MB/s 58 MB/s 34 MB/s
04/11/23 27
Energy Savings
04/11/23 28
State Transitions
04/11/23 29
Response Times
04/11/23 30
Berkeley Web Trace
04/11/23 31
EEVFS Summary
• Knowledge of requests assumed and may be hard to come by
• Performance tied to one of the buffer disks
04/11/23 32
Parallel Striping Groups
Disk 1 Disk 2
Group 1
Buffer Disk
Storage Node 1
Disk 3 Disk 4Buffer Disk
Storage Node 2
Disk 5 Disk 6
Group 2
Buffer Disk
Storage Node 3
Disk 7 Disk 8Buffer Disk
Storage Node 4
File 1 File 2File 3 File 4
04/11/23 33
Striping Within a Group
Disk 1 Disk 2
Group 1
Buffer Disk
Storage Node 1
Disk 3 Disk 4Buffer Disk
Storage Node 2
1 3 5 7 9 4 6 8
4 6 81 3 5 7 9
10
10
1
2
1
2
File 1 File 22 2
04/11/23 34
Striping Within a Group
• Number of disks in a group can be matched to nearest bottleneck
• Striping within the group maintains relatively high performance
• Allows us to use a buffer disk for each storage node, while still maintaining file striping level
04/11/23 35
TestbedParameter Storage Server Storage Node
CPU Celeron 2.2 GHz Celeron 2.2 GHz
Memory (MB) 2000 2000
Network Interconnect
1000 1000
Disk Type SATA SATA
Disk Capacity 160 GB 480 GB
Disk Bandwidth 126 MB/s 126 MB/s
04/11/23 36
Measured Results
04/11/23 37
Measured Results
04/11/23 38
Berkeley Web Trace
04/11/23 39
Response Time Comparison
• Energy efficiency is slightly improved
• Response time gain is significant
Parameter Striping No Striping
Energy Consumption (J) 2,088,113 2,100,243
Response Time (S) 2.78 13.87
04/11/23 40
Parallel Striping Groups Summary
• Improves the energy efficiency and performance of a storage system
• Designed to scale– Needs to be tested on large scale storage
system
04/11/23 41
Conclusions
• Modeling and simulation used to test our ideas
– System, Disk, Trace Parameters varied to study their impacts
• DiskSim Modifications
– Added disk power models to DiskSim
– Implemented block to file translator
• Energy Aware Virtual Cluster File System (EEVFS)
– Implemented a prototype
– Added parallel striping groups to improve the energy efficiency
04/11/23 42
Future Work
• Improve the EEVFS prototype for production use
• Run EEVFS on large scale storage system– Investigate scaling effects
04/11/23 43
http://www.auburn.edu/~xzq0001
Download the presentation slides
Download the presentation slides
Download the presentation slides
Questions