Delay Analysis of Multi-Priority Computing Tasks in Alibaba Cluster Traces

Chenyu Gong, Mulei Ma, Liekang Zeng, Yang Yang, Xiaohu Ge, Liantao Wu

Abstract: Future 6G network AI architecture will provide customized services for everyone, so it is necessary to concentrate on user requirements. Most applications adopt delay as a critical indicator, but few studies have explored its distribution's underlying causes in detail. To address this issue, we analyzed some comparable datasets, i.e., Alibaba Cluster Traces for the years 2017 and 2018, aiming to get some insights from the facts. Traces contain both high-priority online tasks and low-priority batch tasks. We observed the distributions of multiple fields and found several insights: (i) Due to the low priority, the planned CPU cores of batch tasks will be limited by the online tasks. In other words, the service delay of batch tasks will be impacted by the high-priority online tasks. (ii) The overall distribution of task delays and planned CPU cores has upper and lower bounds, both of which can be fitted with logarithmic functions. (iii) The future network AI architecture will be multi-tier and resource fine-grained. Because the flexibility of resource allocation for batch tasks will increase as the CPU granularity of the online tasks grows.