Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris...
Transcript of Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris...
![Page 1: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/1.jpg)
Models and Metrics forEnergy-Efficient Computer Systems
Suzanne RivoireMay 22, 2007Ph.D. DefenseEE Department, Stanford University
![Page 2: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/2.jpg)
Power and Energy Concerns Processors: power density
[Borkar, Intel]
![Page 3: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/3.jpg)
Power and Energy Concerns (2) Personal computers
Mobile devices: battery life/usability Desktops: electricity costs, noise
Servers and data centers Power and cooling costs Reliability Density/scalability Pollution Load on utilities
![Page 4: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/4.jpg)
Underlying Questions Metrics: What are we aiming for?
Compare energy efficiency Identify / motivate new designs
Models: How do we get there? Understand how high-level properties affect
power Improve power-aware scheduling policies /
usage
![Page 5: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/5.jpg)
Talk Overview
Metrics: JouleSort benchmark First complete, full-system energy-efficiency
benchmark Design of winning system
Models: Mantis approach Generates family of high-level full-system
models Generic, accurate, portable
![Page 6: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/6.jpg)
JouleSort energy-efficiency benchmark
JouleSort benchmark specification Workload, metric, guidelines Rationale and pitfalls
Energy-efficient system design:2007 “winner” 3.5× better than previous best Insights for future designs
[S. Rivoire, M. A. Shah, P. Ranganathan, C. Kozyrakis, “JouleSort:A Balanced Energy-Efficiency Benchmark,” SIGMOD 2007.]
![Page 7: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/7.jpg)
Why a benchmark?
Track progress, compare systems, spurinnovation
Current benchmarks/metrics
Limitations of current metrics: Under-specified or “under construction” Limited to a particular component or domain
![Page 8: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/8.jpg)
Benchmark design goals
Holistic and balanced: exercises all corecomponents
Inclusive and representative: meaningfuland implementable on many differentmachines
History-proof: meaningful comparisonsbetween scores from different years
![Page 9: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/9.jpg)
Benchmark specification overview
Workload
Metric
Rules
![Page 10: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/10.jpg)
Workload: External sort
Sort randomly permuted 100-byte recordswith 10-byte keys
From file on non-volatile store to file onnon-volatile store (“external” storage)
![Page 11: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/11.jpg)
External sort workload
Simple and balanced Exercises all core components CPU, memory, disk, I/O, OS, filesystem End-to-end measure of improvement
Inclusive of variety of systems PDAs, laptops, desktops, supercomputers
Representative of sequential I/O tasks Technology trend bellwether Supercomputers to clusters, GPU?
![Page 12: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/12.jpg)
Existing sort benchmarks Sort benchmarks used since 1985
Pure performance MinuteSort: How many records sorted in 1 min? Terabyte: How much time to sort 1 TB?
Price-performance PennySort: How many records sorted for $0.01? Performance-Price: MinuteSort/$$
More info at http://research.microsoft.com/barc/SortBenchmark/
![Page 13: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/13.jpg)
JouleSort metric choices How to weigh power and performance?
Equally (energy)? Energy (Joules) = Power (Watts) × Time (sec.)
Privilege performance (energy-delay product)?
What to fix and what to compare? Fix energy budget and compare records sorted? Fix num. records and compare energy? Fix time budget and compare records/Joule?
![Page 14: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/14.jpg)
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
1.0E+05 1.0E+06 1.0E+07 1.0E+08 1.0E+09 1.0E+10
Records Sorted
SR
ecs/J
.
Records Sorted
Problem with Fixed Time Budget1-pass sort
< 10 sec (N lg N)complexity
Sor
tedR
ecs/
Joul
e
![Page 15: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/15.jpg)
Final metric: Fixed input size
3 classes: 10GB, 100GB, 1TB Winner: minimum energy Report (records sorted / Joule)
Inter-class comparisons imperfect Adjust classes as technology improves
![Page 16: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/16.jpg)
Energy measurement setup
Power
Powermeter
Sorting system
Monitoring system
Wall AC power
Power readings(serial cable) Sort timing
(network)
![Page 17: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/17.jpg)
Talk Overview
Metrics: JouleSort benchmark First complete, full-system energy-efficiency
benchmark Design of winning system
Models: Mantis approach Generates family of high-level full-system
models Generic, accurate, portable
![Page 18: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/18.jpg)
Representative systems
406
22
140
90
290
Pwr (W)~320059GBn/a9GPUTeraSort
(estimated)
~340010GB1%1Laptop
~380010GB>90%12Commodityfileserver
~120010GB26%2Low-endserver
11%
CPU %
~3005GB1Blade
SRecs/JSRecsDisks
![Page 19: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/19.jpg)
Representative systems
406
22
140
90
290
Pwr (W)~320059GBn/a9GPUTeraSort
(estimated)
~340010GB1%1Laptop
~380010GB>90%12Commodityfileserver
~120010GB26%2Low-endserver
11%
CPU %
~3005GB1Blade
SRecs/JSRecsDisks
![Page 20: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/20.jpg)
Representative systems
406
22
140
90
290
Pwr (W)~320059GBn/a9GPUTeraSort
(estimated)
~340010GB1%1Laptop
~380010GB>90%12Commodityfileserver
~120010GB26%2Low-endserver
11%
CPU %
~3005GB1Blade
SRecs/JSRecsDisks
![Page 21: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/21.jpg)
Energy-Efficient Components:Processor
52% power
75% perf
Fileserver CoolSort
Sort BW: 313 MB/s
65W (peak)
Sort BW: 236 MB/s
34W (peak)
![Page 22: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/22.jpg)
Energy-Efficient Components: Disks
15% power
50% perf
Fileserver Our winner
Seagate BarracudaSeq. BW: 80MB/s
13W
Hitachi TravelstarSeq. BW: 40MB/s
2W
![Page 23: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/23.jpg)
CoolSort DesignAsus motherboard:
Mobile CPU + 2 PCI-e slots
RocketRAID Disk Controllers
13 Hitachi TravelStar 160GB
![Page 24: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/24.jpg)
0
2000
4000
6000
8000
10000
12000
2 3 4 5 6 7 8 9 10 11 12 13
Disks Used
Sort
edR
ecs/
Joul
e
0
20
40
60
80
100
120
140
Sort
edR
ecs/
sec
(x 1
0E4)
SRecs/J Perf
Maximizing performance Balanced sort: enough disks to fully utilize CPU Disks running near peak BW
GPUTeraSort
![Page 25: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/25.jpg)
CoolSort: The 100 GB winner
11,300 records sorted per Joule
3.5× more efficient than GPUTeraSort
Average sorting power: 100 W
![Page 26: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/26.jpg)
Insights for future designs Low-hanging fruit: use low-power HW
Best power-performance trade-off Still need to fully utilize resources Challenge: adequate interfaces and “glue” to bring laptop
components into servers
Scaledown efficiency Limited dynamic range For fixed HW: peak efficiency = peak performance How can we design machines that perform equally well in
different benchmark classes?
![Page 27: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/27.jpg)
Benchmark limitations
Tests energy efficiency at high utilization --but most servers are under-utilized How efficient is system at 50% utilization?
20%?
Doesn’t measure building power/cooling
Real goal: TCOSort JouleSort and PennySort give pieces of the
answer
![Page 28: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/28.jpg)
JouleSort Conclusions Need energy-efficiency benchmark JouleSort specification
Simple, representative, full-system benchmark Workload, metric, measurement rules
CoolSort system 3.5× better than 2006 estimated winner Mobile components, server-class interfaces
Part of the sort benchmark suite joulesort.stanford.edu
![Page 29: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/29.jpg)
Talk Overview
Metrics: JouleSort benchmark First complete, full-system energy-efficiency
benchmark Design of winning system
Models: Mantis approach Generates family of high-level full-system
models Generic, accurate, portable
![Page 30: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/30.jpg)
Who needs power models?
Component and system designers How do design decisions affect power?
Users How do my usage patterns affect power?
Data center schedulers How will workload distribution decisions affect
power?
![Page 31: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/31.jpg)
Power modeling goals
Goal: Online, full-system power models Model requirements
Non-intrusive and low-overhead Easy to develop and use Fast enough for online use Reasonably accurate (within 10%) Inexpensive Generic and portable
![Page 32: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/32.jpg)
Power modeling approaches
Detailed component models Simulation-based Hardware metric-based
High-level full-system models
![Page 33: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/33.jpg)
Detailed models: Simulation-based
Inexpensive, arbitrarily accurate Not full-system Slow (not real-time) Not portable
Input:- Current state- Architecture- Circuit parameters
SimulationOutput:Predicted power(component)
![Page 34: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/34.jpg)
Detailed models: Metric-based
Highly accurate Not full-system Complex, require specialized knowledge Not portable
Input:- Design info- HW counters
EquationOutput:Predicted power(component)
[Contreras and Martonosi, ISLPED 2005][Isci and Martonosi, MICRO 2003]
![Page 35: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/35.jpg)
High-level metrics (Mantis)
How accurate? How portable? Tradeoff between model parameters/complexity
and accuracy?
Input:Common util.metrics
EquationOutput:Predicted power(system)
![Page 36: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/36.jpg)
Power Modeling Run one-time calibration scheme
(possibly at vendor) Stress individual components: CPU,
memory, disk Outputs: time-stamped performance
metrics & AC power measurements Fit model parameters to calibration
data Use model to predict power
Inputs: performance metrics at eachtime t
Output: estimation of AC power ateach time t
![Page 37: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/37.jpg)
Models studied
Constant power (the null model):
CPU utilization-based models
!
P = C0
Input:CPU util. %
EquationOutput:Predicted power(system)
![Page 38: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/38.jpg)
CPU utilization-based models
Linear in CPU utilization
Empirical power model
[Fan et al, ISCA 2007]
!
P = C0
+ C1u + C
2u
r
!
P = C0
+ C1u
![Page 39: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/39.jpg)
CPU + disk utilization
Input:- CPU util. %- Disk util. %
EquationOutput:Predicted power(system)
!
P = C0
+ C1u
CPU+ C
2u
disk
[Heath et al, PPoPP 2005]
![Page 40: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/40.jpg)
CPU + disk util. + performance ctrs
Input:- CPU util. %- Disk util. %- CPU perfctrs
EquationOutput:Predicted power(system)
!
P = C0
+ C1u
CPU+ C
2u
disk+ C
iP
i"
[D. Economou, S. Rivoire, C. Kozyrakis,P. Ranganathan, MoBS 2006]
![Page 41: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/41.jpg)
CPU performance counters Configurable processor registers to count
microarchitectural events Requires OS modification In this study:
Memory bus transactions Unhalted CPU clock cycles Instructions retired/ILP Last-level cache references Floating-point instructions
![Page 42: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/42.jpg)
Evaluation methodology
Run calibration suite and develop modelson a variety of machines
Run benchmarks, collecting metrics andAC power
Compare predicted power from metricswith measured AC power
![Page 43: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/43.jpg)
Evaluation machines
CoolSort with 1 and 13 disks Highest and lowest frequencies
2005-era AMD laptop Highest and lowest frequencies
2005-era Itanium server 2008-era Xeon server with 32 GB FBDIMM Variety in component balance, processor,
domain, dynamic range
![Page 44: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/44.jpg)
Evaluation benchmarks
SPECcpu int and fp Laptop: gcc and gromacs only
SPECjbb Stream I/O-intensive programs
ClamAV Nsort (CoolSort-13 only) SPECweb (Itanium only)
![Page 45: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/45.jpg)
Overall mean % error
![Page 46: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/46.jpg)
Overall mean % error
Any model is more accurate than none, andmore detail/complexity is better than less.
![Page 47: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/47.jpg)
Overall mean % error
Performance counter model is most accurateacross the board.
Any model is more accurate than none, andmore detail/complexity is better than less.
![Page 48: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/48.jpg)
Overall mean % error
Performance counter model is most accurateacross the board.
Any model is more accurate than none, andmore detail/complexity is better than less.
Simple linear CPU-util. model gets within10%…with some exceptions.
![Page 49: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/49.jpg)
Best case for empirical CPU model(Xeon server)
![Page 50: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/50.jpg)
Best case for empirical CPU model(Xeon server)
Useful to model shared resources andbottlenecks
![Page 51: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/51.jpg)
Best case for performance counters(Xeon server and CoolSort-13)
![Page 52: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/52.jpg)
Best case for performance counters(Xeon server and CoolSort-13)
Necessary when dynamic memory power is high
![Page 53: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/53.jpg)
Best case for performance counters(Xeon server and CoolSort-13)
Necessary when dynamic memory power is high
Useful to tell how CPU is being utilized
![Page 54: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/54.jpg)
Modeling conclusions Generic approach to power modeling yields
accurate results Simple models overall have < 10% error Same parameters across very different machines More information better models
Linear CPU util. model not enough for… Machines and workloads that are not CPU-dominated CPUs with shared resource bottlenecks Aggressively power-optimized CPUs …all of which reflect hardware trends.
![Page 55: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/55.jpg)
Future work
Beyond CPU, memory, and disk GPUs Network (not a factor today)
Model complexity Combine exponential CPU model w/ perfctrs? Cooling?
![Page 56: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/56.jpg)
Overall Summary Models and metrics needed to improve
energy efficiency Metrics:
JouleSort energy-efficiency benchmarkspecification
Winning JouleSort machine Models:
Simple, portable high-level modeling technique Trade-offs between accuracy and simplicity
![Page 57: Models and Metrics for Energy-Efficient Computer Systems · Collaborators: Mehul Shah, Dimitris Economou, Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie Orgish, Teresa Lynn](https://reader034.fdocuments.in/reader034/viewer/2022050507/5f989d1e2ad10b53283e19db/html5/thumbnails/57.jpg)
Acknowledgments
Advisor: Christos Kozyrakis Mentor: Partha Ranganathan Committee: Kunle Olukotun & Dwight Nishimura Collaborators: Mehul Shah, Dimitris Economou,
Justin Meza Assistance: Jacob Leverich, HP Labs, Charlie
Orgish, Teresa Lynn Defense food! Jayanth and Amin Architecture grad students Grant Gavranovic, Kelley Rivoire, friends & family