Predictive Design Space Exploration Using Genetically Programmed Response Surfaces Henry Cook Department of Electrical Engineering and Computer Science.

Predictive Design Space Exploration Using Genetically

Programmed Response Surfaces

Henry CookDepartment of Electrical Engineering and Computer Science

University of California, Berkeley

[email protected]

Kevin SkadronDepartment of Computer Science

University of Virginia

[email protected]

Designs Are Growing More Complex

More transistors = more complex designs Superscalar, out-of-order execution Chip multiprocessors, systems-on-a-chip Heterogeneity, shared memory, interconnects…

Many interrelated design variables with emergent interactions Unclear whether analytic models can be

constructed based on a priori understanding

How can we find the best design?

Detailed, cycle-accurate simulations Testing every option is time consuming

1 min. x 1 billion simulations = too long! Even search requires too many

simulations Solution: Predict which designs are best

based on a small data sample

Why not heuristic search?

Every step of search requires at least one full simulationHow many steps?

Predicting based on a predetermined sample becomes increasingly efficient as space growsAssuming predictions are accurateAssuming sample is small enough

How can we address complexity?

We need a tool to predict global performance based on a data sample

Many techniques might do this, but we want to use one that… Is more accurate when using small samples Finds the true optimal solutions Gives the architect more insight

Not just ‘what’ the best answer is, but also ‘why’ it is the best

How can we make predictions?

Build a ‘response surface’ A function that relates

design choices to a performance measure

Predicts performance for designs we did not ever simulate

Must have high accuracy to be useful

A simple response surface

Linear Regression:

A simple response surface

Linear Regression:

A simple response surface

Linear Regression:

A simple response surface

Non-linear models are more precise:

How to use response surfaces

The true performance data:

How to use response surfaces

Step 1: Run simulations to make a sample

How to use response surfaces

Step 2: Build the response surface

How to use response surfaces

Step 3: Predict which designs are optimal

How can we build the surface?

Linear/polynomial regression Artificial neural networks Genetic Programming

AutomatedRobustInsightful

Genetic Programming

Creates non-linear, polynomial functions which match sample data

Previous uses Calcination of cement Stress fractures in steel Branch prediction strategy

Evolutionary algorithm Natural selection Survival of the fittest

Evaluate fitness of new individuals

Recombine best individuals

Remove lowest quality individuals

Evolution and reproduction

Result: Explicit response surface functions which best match data

Evolution and reproduction

Expression trees define the structure of the response surface equation

Tuning parameters provide best possible fit to collected data

Candidates are evaluated based on distance of predictions from sampleafter tuning

Combine

Evolution and reproduction

Proving the GP technique

Ipek et. al (ASPLOS-XII) 12 design choices, 20K+ possible designs Predicted IPC and cache performance of 11 applications

Lee and Brooks (ASPLOS-XII) 13 design choices, 1 billion+ possible designs Predicted power and BIPS of 7 applications

Took simulation data from previous studies Designs are realistic but (comparatively) simple

How good are the predictions?

0.1% of possible designs Branch prediction and L2 cache miss rate

• <1% mean error and <2% s.d. Instructions per cycle

• 2.8-4.9% mean error, 2.2-3.4% s.d. 0.000002% of possible designs

Billions of instructions per second• 1.1-6.1% mean, 2.3-9.9% s.d.

Power (W)• 3.5-6.2% mean, 3.4-5.7% s.d.

How good are the predictions?

CDFs of prediction accuracy of BIPS/IPC

How good are the predictions?

Localized accuracy of IPC for Ipek

Can we find optimal designs?

Surface allows us to solve for the best values for some variables analytically Other variables have insignificant impact

When checked by exhaustive search, for most benchmarks 100% of the optimal designs had the analytically determined values

Worst design with those values still had performance that was 97% of optimal

How does the technique help architects?

Saves time Creates response function automatically Only need to simulate predetermined sample points High accuracy means confidence in predictions

Provides insight Which design choices are important What the correct choices are

Limitations

Time to construct response surface via GPRS rather than with ANNs Hours rather than minutes on this space

• Still just an up front cost on the order of a single simulation

Feedback is an explicit, analytic expression How to determine quality without exhaustive

search? Cross validation

Questions?

Thanks to John Lach, Paul Reynolds, Sally McKee, David Brooks, Karan Singh, Benjamin Lee