Unleashing the Power of GPUs over the Web Vishal Vaidyanathan - PowerPoint PPT Presentation

Unleashing the Power of GPUs over the Web Vishal Vaidyanathan Royal Caliber LLC

GPUs are great... ● high compute throughput ● increasingly energy efficient ● high density … but? http://onu.io vishal@onutechnology.com Onu Technology, Inc.

… but come at a price ● high development time ● unknown outcome for many algorithms ● requires specialized talent ● increased operational complexity http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Scalable computing is hard! ● parallel algorithms are already challenging ● CPUs and GPUs are far from ideal parallel processors and interconnects lag far behind How can we bring scalable computation to the mainstream? http://onu.io vishal@onutechnology.com Onu Technology, Inc.

To the cloud! ● API for computation exposed over HTTP ● Powered by hybrid CPU/GPU clusters ● Built-in performance and scalability http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Cloud accessibility http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Broad suite of algorithms Composable blocks Ready-to-use ● basic transforms: sorting, ● fully automated prediction filtering, sampling, reduction, ● image recognition etc. ● graph routing ● specializations: images, matrices, graphs, video, audio, etc. ● machine learning: regression, classification, deep learning, NLP etc. Royal Caliber LLC

System Overview HTTP Browser UI Scheduler API Service High Performance Compute Backend SDK HTTP Thin wrappers in python, java, ruby, scala etc. http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Functional Separation How API Service HTTP Parallelism, What performance, application domain scalability http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Built-in Task Parallelism x h f API Service y = f(x) HTTP z = g(y) y w w = h(x) g z http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Streaming, Visual, Fault Tolerant x h f API Service y = f(x) HTTP z = g(y) y w w = h(x) g z http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Demo: Image recognition var api = new onu.API(); var labels = api.labelImage(url).readAsJson() labels = {‘tabby cat’: 0.7, ‘terrier’: 0.1, ...} http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Demo: Graph analytics var graph = api.newVariable(‘/datasets/wikipedia’) var ranks = api.pageRank(graph) var top = api.argsort(ranks).slice(1000) var topRanks = api.gather(ranks, top) var topNames = api.gather(graph.names, top) var dist = api.shortestPaths(graph, top, top) plot(...) http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Demo: Automated Prediction var model = api.autopred(data, {inCols:[0, 1, 2], outCols:[3]}) model.predict([0.5, 0.12, 2.3]) http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Summary: http://onu.io ● HPC delivered as a service ● ready-to-use and scalable ● intuitive API for ML and data science ● language agnostic http://onu.io vishal@onutechnology.com Onu Technology, Inc.

Acknowledgements Onu: Guha J., T. Sahai, Shriphani P., E. Elsen U.C. Davis: Prof. John Owens, Yangzihao Wang DARPA: Dr. Chris White, Brian Sandberg http://onu.io vishal@onutechnology.com Onu Technology, Inc.