High Performance Data Intensive Computing Dongfang Zhao, Assistant - - PowerPoint PPT Presentation

High Performance Data Intensive Computing

Dongfang Zhao, Assistant Professor Department of Computer Science & Engineering University of Nevada, Reno

Who am I

• 2017 – , Assistant Professor, University of Nevada, Reno
• 2016, Postdoctoral Fellow, University of Washington, Seattle
• 2015, PhD in Computer Science, Illinois Institute of Technology, Chicago
• 2015, Summer Intern, IBM Research – Almaden, San Jose, CA
• 2009-2011, Software Engineer, Epic Systems, Madison, WI
• 2008, MS in Computer Science, Emory University, Atlanta, GA
• 2005, MS in Statistics, Katholieke Universiteit Leuven, Belgium

Outline

• Past Work

– 2005-2008: Machine Intelligence, Computer Vision – 2012-2015: High Performance Computing, Distributed Systems – 2015-2016: Big Data Systems, Database Systems

• Current Status

– Personnel – Facilities

• Future Research Directions

– Distributed Memory Management for Big Data Systems – Locality-aware Resource Management in Virtualized Computing – High Performance Database Systems

Past Work: 2005-2008

• Incremental Dimensionality Reduction
• E.g., published at IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI)

Past Work: 2012-2015

• High Performance Computing
• E.g., published at IEEE Transactions on Parallel and Distributed Systems (TPDS)

Past Work: 2015-2016

• Big Data Systems
• E.g., published at Very Large Database Systems (VLDB)

Current Status: Personnel

• Currently at Nevada:

– 1 PhD student starting Fall 2017 – 1 master student starting Fall 2017 – Collaborating closely with Prof. Dr. Feng Yan working on Data Mining and Performance Modelling. He used to publish at KDD, SIGMETRICS, Supercomputing, CLOUD, NOMS, etc.

• Plan:

– By Fall 2018, the lab will recruit: two more PhD students, two more master students

Current Status: Facilities

• Nevada’s HPC cluster

– 56 compute nodes: PowerEdge C6320

• 1792 cores
• 128 (or 192?) GB RAM per node

– 11 GPU Nodes: PowerEdge C4130 each with 4xP100 with NVLink

• 352 cores
• 44 P100 GPUs
• Our lab’s 10-node GPU cluster, each node has

– 12 CPU cores – 4 GeForce GTX 1080 cards – 64 GB RAM

Future Directions

• Distributed Memory Management for Big Data Systems

– Motivation: Modern big data systems do not have a coordinated way to manage memory

• Users are asked to specify the memory allocation
• Local OS takes the responsibility

– Objective

• A middleware to automatically manipulate memory for big data systems
• The middleware oversees the overall memory status rather than optimizing the local usage
• Users should be able to plug in ad-hoc strategy for the underlying memory management

Future Directions

• Locality-aware Resource Management in Virtualized Computing

– Extension of my intern work in Summer 2015 – Motivation: Load balance is sometimes overemphasized – Objective: improve data locality for virtualized computation

Future Directions

• High Performance Distributed Databases

– Motivation: for some reason, HPC’s dominant storage solution is file system – Objective: building a high-performance distributed database system atop existing parallel/distributed file systems that will support performant:

• Queries expressed in SQL
• Data load, transform, extract, etc.

– Challenges

• Performance bottleneck: from network to what?
• How to leverage GPUs, InfiniBand, MPI, etc. for database workloads?
• …

Thanks!

Dongfang Zhao dzhao@unr.edu