Workload Intelligence: Punching Holes Through the Cloud Abstraction

• DOI: 10.48550/arxiv.2404.19143
• Publication Date: 2024-04-29
• Authors: Lexiang Huang; Anjaly Parayil; Jue Zhang; Xiaoting Qin; Chetan Bansal; Jovan Stojkovic; Pantea Zardoshti; Pulkit A. Misra; Eli Cortez; Raphael Ghelman; Íñigo Goiri; Saravan Rajmohan; Jim Kleewein; Rodrigo Fonseca; Timothy Zhu; Ricardo Bianchini
• Keywords: Abstraction

Today, cloud workloads are essentially opaque to the platform. Typically, only information platform receives is virtual machine (VM) type and possibly a decoration (e.g., VM evictable). Similarly, receive little no from platform; generally, might telemetry their VMs or exceptional signals shortly before evicted). The narrow interface between platforms has several drawbacks: (1) surge in types decorations public complicates customer selection; (2) essential workload characteristics low availability requirements, high latency tolerance) often unspecified, hindering customization for optimized resource usage cost savings; (3) may be unaware of potential optimizations lack sufficient time react events. In this paper, we propose framework, called Workload Intelligence (WI), dynamic bi-directional communication Via WI, can programmatically adjust key characteristics, even dynamically adapt behaviors like priorities. other direction, WI allows inform about upcoming events, opportunities optimization, among scenarios. Because drastically simplify its offerings, reduce costs without fear violating any prices customers on average by 48.8%.