实验室博士生吴朝锐撰写的论文《Chrono: Efficient Serverless Analytics with Adaptive Fine-grained Partitioning and Shadow Execution》被《 IEEE Transactions on Cloud Computing》录用。 论文将于2026年正式发表。
Abstract—Serverless computing has been widely adopted in large-scale data analytics. However, a critical challenge in data analytics is the substantial variability in the volume of data assigned to individual function invocations across executions. This high degree of inter-invocation variability introduces performance unpredictability and complicates efficient resource utilization. Existing approaches attempt to address this issue through vertical scaling, but their effectiveness is inherently limited, as function performance quickly saturates with increased resource allocation. To address these limitations, we propose Chrono, an adaptive serverless analytics framework designed to reduce latency and mitigate cold start overhead under high concurrency. Chrono incorporates three key components: (1) continuous monitoring of function execution to proactively detect risks of SLO violations; (2) dynamic partitioning of unprocessed data among parallel execution instances, aiming to reduce the end-to-end latency; and (3) the integration of lightweight shadow functions running in shared environments, effectively reducing cold start latency. Experimental evaluations on real-world datasets indicate that Chrono outperforms existing methods by reducing average latency by up to 43.16% and 45.42% in 99th percentile latency. Regarding resource utilization, Chrono improves multi-core CPU utilization by 60.6%, and decreases maximum memory usage by 27.2%.