Running Serverless Function on Resource Fragments in Data Center

Repetitive scheduling of resources in data centers leads to a large number of transient resource fragments. Serverless computing can decompose long-period mixed deployment applications into small-volume and short-period functions. Fine-grained functions can effectively respond to the transient fluctuation behavior of fragments. However, the fluctuation characteristics and predictability of resource fragments in the cloud environment vary greatly, which may still interfere with the execution process of functions. In this paper, we introduce Tianxuan, a function-level resource scheduler. It classifies the resource fragments in the cloud environment according to their usage forms into unallocated and unused fragments and extracts the fluctuation characteristics from open-source datasets of each type of fragment. Then, it builds a fragment feature classification-aware spatiotemporal prediction model based on these characteristics and schedules functions on spatiotemporally complementary fragments to achieve resource efficiency by reducing the probability of resource contention. Experimental results show that Tianxuan improves the prediction accuracy of resource fragments by 15% compared to existing techniques and increases server CPU utilization by 16%–25% and function throughput by 17%–37%.