In recent years, Kubernetes (K8s) has become a dominant resource management and scheduling system in the cloud. In practical scenarios, short-running cloud workloads are usually scheduled through different scheduling algorithms provided by Kubernetes. For example, artificial intelligence (AI) workloads are scheduled through different Volcano scheduling algorithms, such as $GANG\_MRP$, $GANG\_LRP$, and $GANG\_BRA$. One key challenge is that the selection of scheduling algorithms has considerable impacts on job performance results. However, it takes a prohibitively long time to select the optimal algorithm because applying one algorithm in one single job may take a few minutes to complete. This poses the urgent requirement of a simulator that can quickly evaluate the performance impacts of different algorithms, while also considering scheduling-related factors, such as cluster resources, job structures and scheduler configurations. In this paper, we design and implement a Kubernetes simulator called K8sSim, which incorporates typical Kubernetes and Volcano scheduling algorithms for both generic and AI workloads, and provides an accurate simulation of their scheduling process in real clusters. We use real cluster traces from Alibaba to evaluate the effectiveness of K8sSim, and the evaluation results show that (i) compared to the real cluster, K8sSim can accurately evaluate the performance of different scheduling algorithms with similar $CloseRate$ (a novel metric we define to intuitively show the simulation accuracy), and (ii) it can also quickly obtain the scheduling results of different scheduling algorithms by accelerating the scheduling time by an average of 38.56×.

Modern distributed engines are increasingly deployed to accelerate large-scaled deep learning (DL) training jobs. While the parallelism of distributed workers/nodes promises the scalability, the computation and communication overheads of the underlying iterative solving algorithms, e.g., stochastic gradient decent, unfortunately becomes the bottleneck for distributed ML training jobs. Existing approaches address such limitations by designing more efficient synchronization algorithms and model compressing techniques, but do not adequately address issues relating to processing massive datasets. In this paper, we propose ClipDL which accelerates the deep learning systems by simultaneously decreasing the number of model parameters and reducing the computations on critical data only. The core component of ClipDL is the estimation of critical set based on the observation that large proportions of input data have little influence on model parameter updating in many prevalent DL algorithms. We implement ClipDL on Spark (a popular distributed engine for big data) and BigDL (based on de-factor distributed ML training architecture, parameter server). The exhaustive experiments on real ML applications and datasets show ClipDL accelerates model training process by an average of 2.32 times while only incurring accuracy losses of 1.86%.