EcoLife: Carbon-Aware Serverless Function Scheduling for Sustainable Computing

TL;DR

This work introduces ECOLIFE, the first carbon-aware serverless function scheduler to co-optimize carbon footprint and performance and designs multiple novel extensions to Particle Swarm Optimization in the context of serverless execution environment to achieve high performance while effectively reducing the carbon footprint.

Abstract

This work introduces ECOLIFE, the first carbon-aware serverless function scheduler to co-optimize carbon footprint and performance. ECOLIFE builds on the key insight of intelligently exploiting multi-generation hardware to achieve high performance and lower carbon footprint. ECOLIFE designs multiple novel extensions to Particle Swarm Optimization (PSO) in the context of serverless execution environment to achieve high performance while effectively reducing the carbon footprint.

Figures (14)

• Figure 1: The carbon footprint (carbon footprint during keeping-alive and service) for three serverless functions for different keep-alive periods. The contribution of carbon footprint during the keep-alive period toward the overall carbon footprint is significant, esp. as the keep-alive period increases.
• Figure 2: The serverless functions can incur a lower overall carbon footprint if kept alive and executed on older-generation hardware due to lower keep-alive carbon footprint (e.g., A$_\textsc{Old}$ vs. A$_\textsc{New}$), but they can suffer from performance degradation. However, the impact on performance can be relatively small for some functions with significant savings in carbon footprint (e.g., Graph-BFS on C$_\textsc{Old}$ vs. C$_\textsc{New}$). The keep-alive period is the same and constant (10 minutes) for all cases.
• Figure 3: Trade-off between carbon footprint and service time: A longer keep-alive period on older-generation hardware can potentially reduce both service time and carbon footprint, but the magnitude and feasibility depend on function characteristics and carbon intensity. Case A: keep alive for 15 mins on C$_\textsc{Old}$, receive a warm start (no cold start overhead) but slower execution. Case B: keep alive for 10 mins on C$_{\textsc{New}}$, and suffer from cold start but faster execution time.
• Figure 4: CO$_2$-Opt, Service-Time-Opt and Energy-Opt are far from the Oracle. There is an opportunity to jointly optimize the overall carbon footprint and service time, but it is challenging to exploit. (A$_\textsc{Old}$ vs A$_\textsc{New}$)
• Figure 5: Optimization process for EcoLife's Dynamic PSO (DPSO) -- particles converge to the optimal after movement.