Our analysis of a large public cloud ML training service shows that resources remain unused likely because users statically (over-)allocate resources for their jobs given a desire for predictable performance, and state-of-the-art schedulers do not exploit idle resources lest they slow down some jobs excessively. We consider if an anticipatory scheduler, which schedules based on predictions of future job arrivals and durations, can improve over the state-of-the-art. We find that realizing gains from anticipation requires dealing effectively with prediction errors, and even the best predictors have errors that do not conform to simple models (such as bounded or i.i.d. error). We devise a novel anticipatory scheduler called SIA that is robust to such errors. On real workloads, SIA reduces job latency by an average of 2.83× over the current production scheduler, while reducing the likelihood of job slowdowns by orders of magnitude relative to schedulers that naïvely share resources.

Emerging SmartNIC-based disaggregated NVMe storage has become a promising storage infrastructure due to its competitive IO performance and low cost. These SmartNIC JBOFs are shared among multiple co-resident applications, and there is a need for the platform to ensure fairness, QoS, and high utilization. Unfortunately, given the limited computing capability of the SmartNICs and the non-deterministic nature of NVMe drives, it is challenging to provide such support on today's SmartNIC JBOFs.

We present Astra, a compilation and execution framework that optimizes execution of a deep learning training job. Instead of treating the computation as a generic data flow graph, Astra exploits domain knowledge about deep learning to adopt a custom approach to compiler optimization. The key insight in Astra is to exploit the unique repetitiveness and predictability of a deep learning job, to perform online exploration of the optimization state space in a work-conserving manner while making progress on the training job. This dynamic state space exploration in Astra uses lightweight profiling and indexing of profile data, coupled with several techniques to prune the exploration state space. Effectively, the execution layer custom-wires the infrastructure end-to-end for each job and hardware, while keeping the compiler simple and maintainable. We have implemented Astra in two popular deep learning frameworks, PyTorch and Tensorflow. On state-of-the-art deep learning models, we show that Astra improves end-to-end performance of deep learning training by up to 3x, while approaching the performance of hand-optimized implementations such as cuDNN where available. Astra also significantly outperforms static compilation frameworks such as Tensorflow XLA both in performance and robustness.