As Artificial Intelligence (AI) is becoming ubiquitous in many applications, serverless computing is also emerging as a building block for developing cloud-based AI services. Serverless computing has received much interest because of its simplicity, scalability, and resource efficiency. However, due to the trade-off with resource efficiency, serverless computing suffers from the cold start problem, that is, a latency between a request arrival and function execution. The cold start problem significantly influences the overall response time of workflow that consists of functions because the cold start may occur in every function within the workflow. Function fusion can be one of the solutions to mitigate the cold start latency of a workflow. If two functions are fused into a single function, the cold start of the second function is removed; however, if parallel functions are fused, the workflow response time can be increased because the parallel functions run sequentially even if the cold start latency is reduced. This study presents an approach to mitigate the cold start latency of a workflow using function fusion while considering a parallel run. First, we identify three latencies that affect response time, present a workflow response time model considering the latency, and efficiently find a fusion solution that can optimize the response time on the cold start. Our method shows a response time of 28–86% of the response time of the original workflow in five workflows.

Gradient sparsification is a widely adopted solution for reducing the excessive communication traffic in distributed deep learning. However, most existing gradient sparsifiers have relatively poor scalability because of considerable computational cost of gradient selection and/or increased communication traffic owing to gradient build-up. To address these challenges, we propose a novel gradient sparsification scheme, DEFT, that partitions the gradient selection task into sub tasks and distributes them to workers. DEFT differs from existing sparsifiers, wherein every worker selects gradients among all gradients. Consequently, the computational cost can be reduced as the number of workers increases. Moreover, gradient build-up can be eliminated because DEFT allows workers to select gradients in partitions that are non-intersecting (between workers). Therefore, even if the number of workers increases, the communication traffic can be maintained as per user requirement. To avoid the loss of significance of gradient selection, DEFT selects more gradients in the layers that have a larger gradient norm than the other layers. Because every layer has a different computational load, DEFT allocates layers to workers using a bin-packing algorithm to maintain a balanced load of gradient selection between workers. In our empirical evaluation, DEFT shows a significant improvement in training performance in terms of speed in gradient selection over existing sparsifiers while achieving high convergence performance.

Graph data structures have been used in a wide range of applications including scientific and social network applications. Engineers and scientists analyze graph data to discover knowledge and insights by using various graph algorithms. A breadth-first search (BFS) is one of the fundamental building blocks of complex graph algorithms and its implementation is included in graph libraries for large-scale graph processing. In this paper, we propose a novel direction selection method, SURF (Selecting directions Upon Recent workload of Frontiers) to enhance the performance of BFS on GPU. A direction optimization that selects the proper traversal direction of a BFS execution between the push and pull phases is crucial to the performance as well as for efficient handling of the varying workloads of the frontiers. However, existing works select the direction using condition statements based on predefined thresholds without considering the changing workload state. To solve this drawback, we define several metrics that describe the state of the workload and analyze their impact on the BFS performance. To show that SURF selects the appropriate direction, we implement the direction selection method with a deep neural network model that adopts those metrics as the input features. Experimental results indicate that SURF achieves a higher direction prediction accuracy and reduced execution time in comparison with existing state-of-the-art methods that support a direction-optimizing BFS. SURF yields up to a $5.62\times$ and $3.15\times$ speedup over the state-of-the-art graph processing frameworks Gunrock and Enterprise, respectively.