Automatic performance model construction for the fast software exploration of new hardware designs

Cavazos, John; Dubach, Christophe; Agakov, Felix; Bonilla, Edwin V.; O’Boyle, Michael; Fursin, Grigori; Temam, Olivier

doi:10.1145/1176760.1176765

Cited by 62 publications

(60 citation statements)

References 26 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Another method of dynamically characterizing a program is through reactions. Here, we record the performance impact from a fixed set of different optimization sequences [9,15]. Thus, we run a program K number of times by optimizing it with K different optimization sequences, and then obtain different speedups for each sequence.…”

Section: Different Ways To Characterize the Programmentioning

confidence: 99%

“…This model has been used in recent work [9,13,22] to predict good optimization sequences for various different compilers. We refer to this model as the speedup predictor because it takes as input a program's characterization (P) and an optimization sequence (O), and it outputs the predicted speedup over some baseline for that optimization sequence.…”

Section: Speedup Prediction Modelmentioning

confidence: 99%

“…Performance counters can be used to collect information, such as cache behavior, functional unit usage, and dynamic instruction mix from the running program [10,15,22]. Reactions is a technique to dynamically characterize a program by selecting specific program transformations to apply to a program and using the resulting speedups to characterize a program's behavior [9,27].…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Using graph-based program characterization for predictive modeling

Park

Cavazos

Álvarez

2012

Proceedings of the Tenth International Symposium on Code Generation and Optimization

Self Cite

View full text Add to dashboard Cite

Using machine learning has proven effective at choosing the right set of optimizations for a particular program. For machine learning techniques to be most effective, compiler writers have to develop expressive means of characterizing the program being optimized. The current state-of-the-art techniques for characterizing programs include using a fixed-length feature vector of either source code features extracted during compile time or performance counters collected when running the program. For the problem of identifying optimizations to apply, models constructed using performance counter characterizations of a program have been shown to outperform models constructed using source code features. However, collecting performance counters requires running the program multiple times, and this "dynamic" method of characterizing programs can be specific to inputs of the program. It would be preferable to have a method of characterizing programs that is as expressive as performance counter features, but that is "static" like source code features and therefore does not require running the program.In this paper, we introduce a novel way of characterizing programs using a graph-based characterization, which uses the program's intermediate representation and an adapted learning algorithm to predict good optimization sequences. To evaluate different characterization techniques, we focus on loop-intensive programs and construct prediction models that drive polyhedral optimizations, such as auto-parallelism and various loop transformation.We show that our graph-based characterization technique outperforms three current state-of-the-art characterization techniques found in the literature. By using the sequences predicted to be the best by our graph-based model, we achieved up to 73% of the speedup achievable in our search space for a particular platform, whereas we could only achieve up to 59% by other state-of-the-art techniques we evaluated.

show abstract

Section: Different Ways To Characterize the Programmentioning

confidence: 99%

Section: Speedup Prediction Modelmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Using graph-based program characterization for predictive modeling

Park

Cavazos

Álvarez

2012

Proceedings of the Tenth International Symposium on Code Generation and Optimization

Self Cite

View full text Add to dashboard Cite

show abstract

“…Cavazos et al [7] have shown that it is possible to improve similar program characterizations by identifying and then restricting to optimizations which carry the most information using the mutual information criterion. However, these optimizations do not necessarily perform best, they are the most discriminatory and one may not afford to "test" them in production runs.…”

Section: Building the Matching Distribution Dmentioning

confidence: 99%

Collective Optimization

Fursin

Temam

2009

High Performance Embedded Architectures and Compilers

Self Cite

View full text Add to dashboard Cite

Abstract. Iterative compilation is an efficient approach to optimize programs on rapidly evolving hardware, but it is still only scarcely used in practice due to a necessity to gather a large number of runs often with the same data set and on the same environment in order to test many different optimizations and to select the most appropriate ones. Naturally, in many cases, users cannot afford a training phase, will run each data set once, develop new programs which are not yet known, and may regularly change the environment the programs are run on. In this article, we propose to overcome that practical obstacle using Collective Optimization, where the task of optimizing a program leverages the experience of many other users, rather than being performed in isolation, and often redundantly, by each user. Collective optimization is an unobtrusive approach, where performance information obtained after each run is sent back to a central database, which is then queried for optimizations suggestions, and the program is then recompiled accordingly. We show that it is possible to learn across data sets, programs and architectures in non-dynamic environments using static function cloning and run-time adaptation without even a reference run to compute speedups over the baseline optimization. We also show that it is possible to simultaneously learn and improve performance, since there are no longer two separate training and test phases, as in most studies. We demonstrate that extensively relying on competition among pairs of optimizations (program reaction to optimizations) provides a robust and efficient method for capturing the impact of optimizations, and for reusing this knowledge across data sets, programs and environments. We implemented our approach in GCC and will publicly disseminate it in the near future.

show abstract

“…Various learning approaches are used in iterative optimization [1][10] to explore a large optimization space. Machine learning is used in [6] to build a performance model based on a small number of evaluations. It first tests a small set of sample optimizations on a prior set of benchmarks, then analyzes the results in order to identify characteristic optimizations, based on which some further test runs are carried out on the target program.…”

Section: Related Workmentioning

confidence: 99%

A Cost-Aware Parallel Workload Allocation Approach Based on Machine Learning Techniques

Long

Fursin

Franke

2007

Lecture Notes in Computer Science

Self Cite

View full text Add to dashboard Cite

Abstract. Parallelism is one of the main sources for performance improvement in modern computing environment, but the efficient exploitation of the available parallelism depends on a number of parameters. Determining the optimum number of threads for a given data parallel loop, for example, is a difficult problem and dependent on the specific parallel platform. This paper presents a learning-based approach to parallel workload allocation in a costaware manner. This approach uses static program features to classify programs, before deciding the best workload allocation scheme based on its prior experience with similar programs. Experimental results on 12 Java benchmarks (76 test cases with different workloads in total) show that it can efficiently allocate the parallel workload among Java threads and achieve an efficiency of 86% on average.

show abstract

Automatic performance model construction for the fast software exploration of new hardware designs

Cited by 62 publications

References 26 publications

Using graph-based program characterization for predictive modeling

Using graph-based program characterization for predictive modeling

Collective Optimization

A Cost-Aware Parallel Workload Allocation Approach Based on Machine Learning Techniques

Contact Info

Product

Resources

About