How Fluid Accelerates Cloud‑Native Deep Learning Training

ICDE 2022 Acceptance

The International Conference on Data Engineering (ICDE) is an IEEE flagship conference, ranked alongside SIGMOD and VLDB as one of the three top venues in data management and databases. A paper titled Fluid: Dataset Abstraction and Elastic Acceleration for Cloud‑native Deep Learning Training Jobs was accepted as a full long paper at ICDE 2022.

Problem Statement

Running deep‑learning training workloads on cloud‑native platforms (Kubernetes/Docker) brings high elasticity and low‑cost operation, but it also creates severe I/O bottlenecks: complex data access patterns, difficulty matching GPU I/O demand, and inefficient sharing of cached data across jobs.

Proposed Solution – Fluid

Fluid provides a Fluid Dataset abstraction that hides heterogeneous storage back‑ends and introduces an automatically optimized cache engine that adapts to dataset characteristics. The system can elastically scale cache space during training based on real‑time I/O demand, and it can reorder job scheduling using cross‑job cache semantics to improve overall throughput.

Open‑Source Project Details

Fluid is an open‑source project under the Cloud Native Computing Foundation (CNCF) and is hosted at https://github.com/fluid-cloudnative/fluid. Initiated jointly by Alibaba Cloud’s cloud‑native team and the Computer Science Department of Nanjing University, the project has accumulated over 1,000 pull‑request submissions, released seven versions, and was accepted into CNCF in April 2021, filling a gap in elastic data‑caching orchestration within the Kubernetes ecosystem.

Real‑World Impact

In production, Fluid has helped many users significantly improve AI model training performance while reducing the complexity of managing training data. Alibaba Cloud integrates Fluid’s core ideas into its cloud‑native AI suite delivered via the ACK (Alibaba Cloud Kubernetes) service.

Recognition and Broader Innovation

The paper’s acceptance reflects Alibaba Cloud’s ongoing innovations in container‑based AI workloads, including prior work on serverless image distribution that was accepted at USENIX ATC 2021. In early 2022, Forrester’s Wave report placed Alibaba Cloud in the “Leader” quadrant for public‑cloud container platforms, a first for a Chinese vendor.