On Data Locality in Kubernetes
summary
Built a Tesseract-based benchmark on three 8‑CPU 32GB worker nodes to measure data locality effects, showing up to 10x local vs remote throughput for reads and up to 5x end-to-end speedup when consolidating pods on the data node. Implemented a direct data‑forwarding prototype (Vloggy) that yields up to 60x speedups for some transfer patterns and a cache/proxy using Alluxio mounted via CSI instead of sidecars; production tests raise GPU utilization from near idle to about 75%.
AI-written summary · about
chapters (7)
- 0:00Data locality in Kubernetes
- 4:25Benchmarking ML pipelines
- 9:05Data placement impact on I/O
- 13:15Alluxio for data locality
- 21:19Caching for ML workloads
- 24:31Production results with Alluxio
- 26:14Q&A
related talks
34mOn Data Locality in Kubernetes
05mLightning Talk: How to Bring Data Locality to I/O-Intensive Workloads on Kubernetes
45mEmbracing Big Data Workload in Cloud-Native Environment with Data Locality
25mOvercoming I/O Bottlenecks in LLM Training with Open-Source Distributed...
45mEmbracing Big Data Workload in Cloud-Native Environment with Data Locality
07mData Gravity and Kubernetes: Managing Large-Scale Data Ingest With Minimal Latency
31mThe Future of Data on Kubernetes: From Database Management To AI Foundation
25mMaking the Most Out of Your Hardware Accelerators in a Kubernetes Cluster