Loading presentation...

Present Remotely

Send the link below via email or IM

Copy

Present to your audience

Start remote presentation

  • Invited audience members will follow you as you navigate and present
  • People invited to a presentation do not need a Prezi account
  • This link expires 10 minutes after you close the presentation
  • A maximum of 30 users can follow your presentation
  • Learn more about this feature in our knowledge base article

Do you really want to delete this prezi?

Neither you, nor the coeditors you shared it with will be able to recover it again.

DeleteCancel

Make your likes visible on Facebook?

Connect your Facebook account to Prezi and let your likes appear on your timeline.
You can change this under Settings & Account at any time.

No, thanks

PhD Candidacy Presentation

No description
by

Saman Khoshbakht

on 27 October 2013

Comments (0)

Please log in to add your comment.

Report abuse

Transcript of PhD Candidacy Presentation

What Can the Software
Tell us About Power?

Saman Khoshbakht
Dr. Nikitas Dimopoulos
Supervisor
Higher performance
leads to
higher power consumption
Multi cores can improve power efficiency
maybe not for long
Software doesn't use constant power
SPEC2000-gcc
The goal
Finding indicators for power
in the software
Landaur's principle
nKTln2
HSPICE Simulation of a single 6T SRAM cell
TSMC Library, 90nm, minimal size
0.35 fJ
0.032 fJ
Transition
No transition
Pertains a lower theoretical limit of energy for a certain computation
n=Number of bit changes
K=1.38e-23 J/K
T=Temperature (K)
Relating application memory activity
to processor power
Hypothesis:
The power consumed by the processor in memory cells correlates to the amount of memory bit transitions
Memory bit transitions
Transitions: 3
00101011
X
00101
100
X
(new)
(old)
Example:
C. Moore, “Data Processing in ExaScale-Class Computer Systems,” presented at the The Salishan Conference on High Speed Computing, 2011.
Methodology
No Available detailed power simulators
Real Power Measurement
Memory Simulation
Offline Matching and Analysis
Main Components
Real Power Measurement
System under Test
Data Collection
Processor Current
Voltage Lines
System Configuration:
Dell Optiplex 7010
Scientific Linux r6.3 (init 1)
Intel core i7-3770
(100mv/A)
Power Measurement
Taking advantage of ATX standard dedicated processor power lines

Current (A)
Tektronix A622
V1 (v)
V2 (v)
Data Acquisition Device
NI USB-6009
Labview Signal Express 2012
3 channels
Sample Results for gcc
Motivation
~8W (30%)
Motivation
Cycle accurate
Access and Transition patterns
Comparing the results
Making the power models
Error estimation
Memory Simulation
- MARSSx86
Full System Multi core Emulator
QEmu: Quick Emulator Engine
M
icro-
AR
chitectural and
S
ystem
S
imulator for
x86
-based Systems
Cycle-Accurate Simulator
PTLSim
Out-of-order
Superscalar
x86_64 architecture
200K+ inst/sec
Modified to track and record memory access and transitions activity
Sample Results
Offline Analyzer
A set of developed Matlab tools
(10 Million cycles)
(for "ls" command)
Regionizing
Clustering
Modeling
gcc
equake
Compare and match results
by means of markers
Region
Classifies the regions based on size
Categorizing program phases
Ability to train different models
Model Noise reduction
Human error compensation

Each region has certain attributes
example:
Four Linear models compared
Experimental
Results

SPEC CPU INT2000-gcc
Correlation between transitions and accesses to power
Shown for top 90% regions
SPEC CPU FP2000
ammp
equake
ammp
equake
Motivation
August 2013
Motivation
Records processor power
(voltage and current)
Sampling frequency = 15,000 Samples/sec
Road map
Preliminary
Experiments
Power Measurement
Framework
Simulation
MARSSx86
Modeling
Tools
Summary
More Benchmark
Evaluation
Back Annotation
Speculative
Power model
Software-based
Power
Management
So far:

Framework is implemented

Clear correlation between power and memory activity

The Framework can run most applications and benchmarks

Initial model based on memory activities created
MSR Monitoring Tool
Questions?
Validate the hypothesis
using more benchmarks
Dr. Amirali Baniasadi
Dr. Sudhakar Ganti
Committee Members
Thank you!
example:
ammp
Access
Transition
equake
gcc (4 instances)
Unsupervised
Supervised
Road map
More Benchmark
Evaluation
Back Annotation
Speculative
Power model
Software-based
Power
Management
MSR Monitoring Tool
Validate the hypothesis using more benchmarks
Native
Synthetic

MSR monitoring tool to collect performance counters

Back annotate the traces to code

Correlate the code with software power indicators
example:
gcc (4 instances)
Unsupervised
Supervised
Native
SPEC2K
SPEC2006
PARSEC
...
Targeted
(L1, L2, ...)
Controlled
transitions
Synthetic
Background Process

Monitoring Processor's Performance Counters (or
M
odel-
S
pecific
R
egisters)

Ability to keep track of performance events
Memory Trace
Power Trace
Source Code
Improve power models
using software indicators
Offline (Static)
Real-time (Dynamic)
Developer
Compiler
Operating System
PhD Candidacy Proposal
Memory cells
Internal Registers
Data Cache
[Main Memory]
Full transcript