Efficient mart-aided modeling for microarchitecture design space exploration and performance prediction

Abstract

Computer architects usually evaluate new designs by cycleaccurate processor simulation. This approach provides detailed insight into processor performance, power consumption and complexity. However, only configurations in a subspace can be simulated in practice due to long simulation time and limited resource, leading to suboptimal conclusions which might not be applied in a larger design space. In this paper, we propose an automated performance prediction approach which employs state-of-the-art techniques from experiment design, machine learning and data mining. Our method not only produces highly accurate estimations for unsampled points in the design space, but also provides interpretation tools that help investigators to understand performance bottlenecks. According to our experiments, by sampling only 0.02% of the full design space with about 15 millions points, the median percentage errors, based on 5000 independent test points, range from 0.32% to 3.12% in 12 benchmarks. Even for the worst-case performance, the percentage errors are within 7% for 10 out of 12 benchmarks. In addition, the proposed model can also help architects to find important design parameters and performance bottlenecks.

An adaptive sampling scheme guided by BART—with an application to predict processor performance

“…Several empirical and theoretical studies have shown that samples selected adaptively outperform those obtained from conventional sampling schemes in learning a target function. See, for example, Freund et al (1993), Sung & Niyogi (1995), Saar‐Tsechansky & Provost (2001), and Li, Peng & Ramadass (2008). We propose the following sequential sampling algorithm:…”

“…One of the benefits from BART is that BART produces a MCMC sample from the induced posterior over the sum‐of‐trees model space, which can readily be used to keep track of the uncertainty of prediction and to estimate the correlations among design points. Li, Peng & Ramadass (2008) use MART‐guided design to improve prediction, where they use MART as the predictive model and sequentially sample design points that have the largest predictive variances while the minimum distance among each other is maximized. They have to use resampling method to estimate the predictive variance.…”

“…During the process, we make sure that the simulation converges to a stationary distribution and the simulations are reasonably independent. For the Maximin method, we use the same distance function as that in Li, Peng & Ramadass (2008). We repeat all sampling methods 100 times.…”

“…We also compare BGAS with the adaptive sampling method guided by MART in Li, Peng & Ramadass (2008, denoted by LI). To make the methods comparable, we start LI with 20 Maximin samples and then adaptively choose 10 additional design points at each iteration until a total of 200 samples are selected.…”

“…Moreover, we show in this paper that neither AIC nor BIC is efficient in finding complex interactions. Li, Peng & Ramadass (2008) proposed a MART‐aided adaptive sampling method. Two algorithms were used in their paper: (1) Multiple Additive Regression Trees (MART, see Friedman, 2001) for predictive model building and (2) adaptive sampling technique for selecting the unsampled points, based on which the MART model is most likely to be improved.…”