How to Eliminate the I/O Bottleneck and Continuously Feed the GPU While Training in the...
summary
Describes Alluxio’s new Dora decentralized repo sitting between compute and cloud storage, built as worker + server-registry + client with no single point of failure and deployed on Kubernetes for PyTorch workloads. Implements zero-copy geo-copy transfers, a FUSE SDK local cache and S3/HDFS APIs, supports tens of billions of files and targets 99.99% availability. Claims raise ResNet50 GPU use from ~20–30% to ~80–90%, give 2–20× throughput, 30–50% large-file reads, up to 9× latency cuts and fixes up to 100× re-amplification.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:10What Alexio is
- 3:33Alexio as data layer
- 5:17I/O bottlenecks resolved
- 7:04Benefits over storage
- 9:52Dora architecture
- 10:35Dora motivations
- 13:04Dora components
- 14:30Scalability and fault tolerance
- 16:37AI workload optimization
- 19:03AI stages connection
- 20:27Alexio FUSE SDK
- 21:50Testing Dora with LLMs
- 22:33Community resources
- 23:02Q&A
related talks
24mHow to Eliminate the I-O Bottleneck & Continuously Feed the GPU While Training in the Cloud
28mMaximize GPU Utilization for Model Training
38mExploring Distributed Caching for Faster GPU Training with NVMe, GDS, and RDMA
25mOvercoming I/O Bottlenecks in LLM Training with Open-Source Distributed Caching
05m⚡ Lightning Talk: Using Decentralized File System to Optimize ML/AI Workloads on Kubernetes
15mTransparent GPU Sharing in Container Clouds for Deep Learning Workloads
25mHigh Performance Networking for Distributed DL Training in Production K8s
38mAll You Need to Know to Build Your GPU Machine Learning Cloud [B]