A Survey of the Performance Fluctuation in Serverless Computing

초록

In Serverless (FaaS), significant performance variability increases costs and risks SLO violations. This survey synthesizes six key studies to identify the root causes of this variability, categorizing them into four areas: internal platform policies, shared resource contention, temporal dynamics, and measurement methodology. We find that platform-level factors like memory bin-packing and VM heterogeneity structurally induce variance, while resource contention can inflate tail latency by tens of times. Temporal factors, such as diurnal patterns and cold-start spikes, also systematically impact performance. Critically, we confirm that reliable estimates necessitate varianceaware experimental design. This survey integrates these findings to summarize the structural origins of variability and presents practical recommendations for both experimental and system design.

키워드