Deep Learning (DL) applications are used to solve complex problems efficiently. These applications require complex neural network models composed of millions of parameters and huge amounts of data for proper training. This is only possible by parallelizing the necessary computations by so-called distributed deep learning (DDL) frameworks over many GPUs distributed over multiple nodes of a HPC cluster. These frameworks mostly utilize the compute power of the GPUs and use only a small portion of the available compute power of the CPUs in the nodes for I/O and inter-process communication, leaving many CPU cores idle and unused. The more powerful the base CPU in the cluster nodes, the more compute resources are wasted. In this paper, we investigate how much of this unutilized compute resources could be used for executing other applications without lowering the performance of the DDL frameworks. In our experiments, we executed a noise-generation application, which generates a very-high memory, network or I/O load, in parallel with DDL frameworks, and use HPC profiling and tracing techniques to determine whether and how the generated noise is affecting the performance of the DDL frameworks. Early results indicate that it might be possible to utilize the idle cores for jobs of other users without affecting the performance of the DDL applications in a negative way.

The profound impact of recent developments in artificial intelligence is unquestionable. The applications of deep learning models are everywhere, from advanced natural language processing to highly accurate prediction of extreme weather. Those models have been continuously increasing in complexity, becoming much more powerful than their original versions. In addition, data to train the models is becoming more available as technological infrastructures sense and collect more readings. Consequently, distributed deep learning training is often times necessary to handle intricate models and massive datasets. Running a distributed training strategy on a supercomputer exposes the models to all the considerations of a large-scale machine; reliability is one of them. As supercomputers integrate a colossal number of components, each fabricated on an ever decreasing feature-size, faults are common during execution of programs. A particular type of fault, silent data corruption, is troublesome because the system does not crash and does not immediately give an evident sign of an error. We set out to explore the effects of that type of faults by inspecting how distributed deep learning training strategies cope with bit-flips that affect their internal data structures. We used checkpoint alteration, a technique that permits the study of this phenomenon on different distributed training platforms and with different deep learning frameworks. We evaluated two distributed learning libraries (Distributed Data Parallel and Horovod) and found out Horovod is slightly more resilient to SDCs. However, fault propagation is similar in both cases, and the model is more sensitive to SDCs than the optimizer.

The convergence of artificial intelligence, high-performance computing (HPC), and data science brings unique opportunities for marked advance discoveries and that leverage synergies across scientific domains. Recently, deep learning (DL) models have been successfully applied to a wide spectrum of fields, from social network analysis to climate modeling. Such advances greatly benefit from already available HPC infrastructure, mainly GPU-enabled supercomputers. However, those powerful computing systems are exposed to failures, particularly silent data corruption (SDC) in which bit-flips occur without the program crashing. Consequently, exploring the impact of SDCs in DL models is vital for maintaining progress in many scientific domains. This paper uses a distinctive methodology to inject faults into training phases of DL models. We use checkpoint file alteration to study the effect of having bit-flips in different places of a model and at different moments of the training. Our strategy is general enough to allow the analysis of any combination of DL model and framework—so long as they produce a Hierarchical Data Format 5 checkpoint file. The experimental results confirm that popular DL models are often able to absorb dozens of bit-flips with a minimal impact on accuracy convergence.

Extreme-scale computing systems are required to solve some of the grand challenges in science and technology. From astrophysics to molecular biology, supercomputers are an essential tool to accelerate scientific discovery. However, large computing systems are prone to failures due to their complexity. It is crucial to develop an understanding of how these systems fail to design reliable supercomputing platforms for the future. This paper examines a five-year failure and workload record of a leadership-class supercomputer. To the best of our knowledge, five years represents the vast majority of the lifespan of a supercomputer. This is the first time such analysis is performed on a top 10 modern supercomputer. We performed a failure categorization and found out that: i) most errors are GPUrelated, with roughly 37% of them being double-bit errors on the cards; ii) failures are not evenly spread across the physical machine, with room temperature presumably playing a major role; and iii) software errors of the system bring down several nodes concurrently. Our failure rate analysis unveils that: i) the system consistently degrades, being at least twice as reliable at the beginning, compared to the end of the period; ii) Weibull distribution closely fits the mean-time-between-failure data; and iii) hardware and software errors show a markedly different pattern. Finally, we correlated failure and workload records to reveal that: i) failure and workload records are weakly correlated, except for certain types of failures when segmented by the hours of the day; ii) several categories of failures make jobs crash within the first minutes of execution; and iii) a significant fraction of failed jobs exhaust the requested time with a disregard of when the failure occurred during execution. Index Terms-Fault tolerance, resilience, failure analysis, high performance computing.