Design and Practices of an Elastic Computing Platform for Efficient Resource Utilization

The platform was created to address the growing demand for massive image and video processing, AI computation, and the under‑utilization of idle resources in production data centers, where online services exhibit clear peak‑valley patterns and resource fragmentation.

Key challenges include preventing interference with online service quality (capacity, latency, scheduling delays, and fault rates), handling highly variable and heterogeneous elastic resources, and making those resources easily consumable by downstream workloads.

The technical architecture is divided into three layers: an access layer providing service APIs and image/video handling; a scheduling layer that abstracts diverse resources via a name service, performs load balancing, auto‑scaling, peak‑shaving, and gray‑release; and a node layer that enforces resource isolation, conflict detection, and container monitoring.

To protect online services, the platform employs performance models for CPU‑network mixing, CPI (cycles‑per‑instruction) monitoring to detect instruction‑level latency, priority‑based container CPU shares for scheduling latency, and OOM‑priority mechanisms that pre‑empt low‑priority containers during memory pressure.

Elastic resources are exposed through scenario‑based services (image compression, video transcoding, AI inference, log processing) with differentiated SLAs, a CL5 naming service that hides resource heterogeneity, and higher‑level interfaces such as cloud functions that abstract away underlying resource management.

Practical lessons highlight the trade‑off between providing mechanisms versus policies, the necessity of load balancing before scaling, and the importance of choosing simple, widely‑adopted low‑level technologies to reduce failure‑repair costs.