Modeling in Computer Architecture Matthew Jacob. Architecture Evaluation Challenges Skadron,...
-
Upload
laura-phillips -
Category
Documents
-
view
213 -
download
1
Transcript of Modeling in Computer Architecture Matthew Jacob. Architecture Evaluation Challenges Skadron,...
Modeling in Computer Architecture
Matthew Jacob
Architecture Evaluation Challenges• Skadron, Martonosi, August, Hill, Lilja and
Pai, IEEE Computer, Aug 2003
• “Quantitative evaluation is the mainstay, but system complexity makes it troublesome”
• There has been a dramatic shift towards simulation
Simulation is the Preferred Tool
Simulation: Are there any real alternatives?
• “Knee-jerk negative reactions from program committee members … effectively discourages the research community from exploring other useful and possibly more informative modeling techniques”
• “Developing scientific methods for abstracting evaluations to explore large design spaces is imperative”
What about analytical models?
• Example: Karkhanis and Smith, A First-order Model of Superscalar Processors, 31st ISCA 2004– Analytical model for estimating superscalar
processor program CPI (Cycles per Instruction)
What about analytical models?
• Example: Karkhanis and Smith, A First-order Model of Superscalar Processors, 31st ISCA 2004– Analytical model for estimating superscalar
processor program CPI
– 5.8% average error
dcachemissicachemissbrmispesteadystat CPICPICPICPICPI
- Uses “expert knowledge”
How reliable is expert knowledge?• “I think there is a world market for maybe five
computers.” (1943)– Thomas Watson, Chairman, IBM
• “640K ought to be enough for anybody.” (1981)– Bill Gates
• “$100 million dollars is way too much to pay for Microsoft.” (1982)– IBM
• “There is no reason anyone would want a computer in their home.” (1977)– Ken Olson, President, Chairman and Founder, DEC
What is Empirical Modeling?
• Extracting models from measured data– We can use simulators to generate the data
Prediction accuracy
Ease
of I
nter
pret
ation
Linear models
Neural nets
Modeling Out-of-order Superscalars
1. Build models to help understand the relative importance of design parameters and also of their interactions
The first (and still only) systematic approach available
2. Build an accurate predictive modelThe first (and still most efficient) predictive
modeling technique available
3. Demonstrate the use of such models
(P. J. Joseph, Kapil Vaswani)
• `Construction and Use of Linear Regression Models for Processor Performance Analysis’, with P. J. Joseph, Kapil Vaswani, HPCA-12, 2006
• `A Predictive Perfomance Model for Superscalar Processors’, with P. J. Joseph, Kapil Vaswani, MICRO-39, 2006
• `Microarchitecture Sensitive Empirical Models for Compiler Optimizations’, with Kapil Vaswani, P. J. Joseph, Y. N. Srikant, CGO-5, 2007
References