BLAS-on-flash: An Efficient Alternative for Large Scale ML Training and Inference?
summary
Presents a library that implements BLAS and sparseBLAS directly on SSD-resident (flash) matrices so multi-threaded code can operate on hundreds of GB without redesigning for distributed systems. Shows that this BLAS-on-flash approach delivers near in-memory BLAS performance and can support complex algorithms (e.g., eigensolvers) and large-scale ML workflows—demonstrated on topic-model training and extreme multi-label inference—outperforming in-memory ARPACK and Spark in the reported cases.
AI-written summary · about
chapters (19)
- 0:00Start
- 0:26Scope | Memory-Intensive non-DL workloads
- 1:31Distributed ML | Current Landscape
- 2:19"Scalability" | Compact systems
- 6:20gemm Chain View
- 7:01gemm Kernel - Task Creation
- 8:06Access Specifier | Block Definition
- 8:35Sector Alignment | Correct vs fast
- 9:16Software Stack | Architecture II
- 11:19Evaluation Hardware Specifications
- 12:15Evaluation Datasets
- 12:47Evaluation Metrics
- 14:49SVD Choosing the right algorithm
- 16:01Eigenvalues Text datasets
- 17:01Eigensolvers | Cost-effectiveness
- 17:26ISLE | Web-Scale Topic Modeling
- 18:03ISLE Larger models, lower costs?
- 18:30XML Web-scale classification
- 20:22Conclusion
related talks
12mBehemoth: A Flash-centric Training Accelerator for Extreme-scale DNNs
12mFlashNeuron: SSD-Enabled Large-Batch Training of Very Deep Neural Networks
01h 00mThe Blaze High Performance Math Library
26mEfficiently serving LLMs at scale.
17mMachine Learning at the Speed of Light with Spark
33mComputing recommendations at extreme scale with Apache Flink
19mOptimizations and abstractions for sparse machine learning
16mAmeet Talwalkar, Assistant Professor of Computer Science, UCLA @ MLconf SF Part 2