Andrew Lumsdaine

Eric Silk, Swarnita Chakraborty, Nairanjana Dasgupta et al.

Training deep neural networks (DNNs) used in modern machine learning is computationally expensive. Machine learning scientists, therefore, rely on stochastic first-order methods for training, coupled with significant hand-tuning, to obtain good performance. To better understand performance variability of different stochastic algorithms, including second-order methods, we conduct an empirical study that treats performance as a response variable across multiple training sessions of the same model. Using 2-factor Analysis of Variance (ANOVA) with interactions, we show that batch size used during training has a statistically significant effect on the peak accuracy of the methods, and that full batch largely performed the worst. In addition, we found that second-order optimizers (SOOs) generally exhibited significantly lower variance at specific batch sizes, suggesting they may require less hyperparameter tuning, leading to a reduced overall time to solution for model training.

Jesun Sahariar Firoz, Thejaka Amila Kanewala, Marcin Zalewski et al.

The increasing complexity of the software/hardware stack of modern supercomputers results in explosion of parameters. The performance analysis becomes a truly experimental science, even more challenging in the presence of massive irregularity and data dependency. We analyze how the existing body of research handles the experimental aspect in the context of distributed graph algorithms (DGAs). We distinguish algorithm-level contributions, often prioritized by authors, from runtime-level concerns that are harder to place. We show that the runtime is such an integral part of DGAs that experimental results are difficult to interpret and extrapolate without understanding the properties of the runtime used. We argue that in order to gain understanding about the impact of runtimes, more information needs to be gathered. To begin this process, we provide an initial set of recommendations for describing DGA results based on our analysis of the current state of the field.

Michael Hansen, Robert L. Goldstone, Andrew Lumsdaine

What factors impact the comprehensibility of code? Previous research suggests that expectation-congruent programs should take less time to understand and be less prone to errors. We present an experiment in which participants with programming experience predict the exact output of ten small Python programs. We use subtle differences between program versions to demonstrate that seemingly insignificant notational changes can have profound effects on correctness and response times. Our results show that experience increases performance in most cases, but may hurt performance significantly when underlying assumptions about related code statements are violated.