Explaining Deep Learning-Based Networked Systems

While deep learning (DL)-based networked systems have shown great potential in various applications, a key drawback is that Deep Neural Networks (DNNs) in DL are blackboxes and nontransparent for network operators. The lack of interpretability makes DL-based networked systems challenging to operate and troubleshoot, which further prevents DL-based networked systems from deploying in practice. In this paper, we propose TranSys, a novel framework to explain DL-based networked systems for practical deployment. Transys categorizes current DL-based networked systems and introduces different explanation methods based on decision tree and hypergraph to effectively explain DL-based networked systems. TranSys can explain the DNN policies in the form of decision trees and highlight critical components based on analysis over hypergraph. We evaluate TranSys over several typical DL-based networked systems and demonstrate that Transys can provide human-readable explanations for network operators. We also present three use cases of Transys, which could (i) help network operators troubleshoot DL-based networked systems, (ii) improve the decision latency and resource consumption of DL-based networked systems by ~10x on different metrics, and (iii) provide suggestions on daily operations for network operators when incidences occur.

Learning Scheduling Algorithms for Data Processing Clusters

Efficiently scheduling data processing jobs on distributed compute clusters requires complex algorithms. Current systems, however, use simple generalized heuristics and ignore workload structure, since developing and tuning a bespoke heuristic for each workload is infeasible. In this paper, we show that modern machine learning techniques can generate highly-efficient policies automatically. Decima uses reinforcement learning (RL) and neural networks to learn workload-specific scheduling algorithms without any human instruction beyond specifying a high-level objective such as minimizing average job completion time. Off-the-shelf RL techniques, however, cannot handle the complexity and scale of the scheduling problem. To build Decima, we had to develop new representations for jobs' dependency graphs, design scalable RL models, and invent new RL training methods for continuous job arrivals. Our prototype integration with Spark on a 25-node cluster shows that Decima outperforms several heuristics, including hand-tuned ones, by at least 21%. Further experiments with an industrial production workload trace demonstrate that Decima delivers up to a 17% reduction in average job completion time and scales to large clusters.