Pratap Pattnaik
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Pratap Pattnaik has a total of 20 co-authors. Here are the top co-authors by number of shared publications:
- Hao Yu
IBM T. J. Watson Research Center Yorktown Heights NY USA
2 shared publications - Jeremy Kepner
Massachusetts Institute of Technology
2 shared publications - Joefon Jann
IBM Thomas J. Watson Research Center
2 shared publications - Jose E. Moreira
IBM Thomas J. Watson Research Center, Yorktown Heights, NY, USA
2 shared publications - Manoj Kumar
IBM Thomas J. Watson Research Center
2 shared publications - Mauricio Serrano
IBM Thomas J. Watson Research Center
2 shared publications
Latest Publications
- 1 July 2018
- Published by IBM in IBM Journal of Research and Development
- Vol. 62 (4/5) , 10:1-10:12
IBM POWER9 and cognitive computing
- 1 July 2018
- Published by IBM in IBM Journal of Research and Development
- Vol. 62 (4/5) , 10:1-10:12
IBM POWER9 and cognitive computing
- Open Access
Enabling Massive Deep Neural Networks with the GraphBLAS
Deep Neural Networks (DNNs) have emerged as a core tool for machine learning. The computations performed during DNN training and inference are dominated by operations on the weight matrices describing the DNN. As DNNs incorporate more stages and more nodes per stage, these weight matrices may be required to be sparse because of memory limitations. The GraphBLAS.org math library standard was developed to provide high performance manipulation of sparse weight matrices and input/output vectors. For sufficiently sparse matrices, a sparse matrix library requires significantly less memory than the corresponding dense matrix implementation. This paper provides a brief description of the mathematics underlying the GraphBLAS. In addition, the equations of a typical DNN are rewritten in a form designed to use the GraphBLAS. An implementation of the DNN is given using a preliminary GraphBLAS C library. The performance of the GraphBLAS implementation is measured relative to a standard dense linear algebra library implementation. For various sizes of DNN weight matrices, it is shown that the GraphBLAS sparse implementation outperforms a BLAS dense implementation as the weight matrix becomes sparser.
Open AccessEnabling Massive Deep Neural Networks with the GraphBLAS
Deep Neural Networks (DNNs) have emerged as a core tool for machine learning. The computations performed during DNN training and inference are dominated by operations on the weight matrices describing the DNN. As DNNs incorporate more stages and more nodes per stage, these weight matrices may be required to be sparse because of memory limitations. The GraphBLAS.org math library standard was developed to provide high performance manipulation of sparse weight matrices and input/output vectors. For sufficiently sparse matrices, a sparse matrix library requires significantly less memory than the corresponding dense matrix implementation. This paper provides a brief description of the mathematics underlying the GraphBLAS. In addition, the equations of a typical DNN are rewritten in a form designed to use the GraphBLAS. An implementation of the DNN is given using a preliminary GraphBLAS C library. The performance of the GraphBLAS implementation is measured relative to a standard dense linear algebra library implementation. For various sizes of DNN weight matrices, it is shown that the GraphBLAS sparse implementation outperforms a BLAS dense implementation as the weight matrix becomes sparser.
- 28 January 2017
- Published by Springer Nature in International Journal of Parallel Programming
- Vol. 46 (2) , 412-440
Graph Programming Interface (GPI): A Linear Algebra Programming Model for Large Scale Graph Computations
- 28 January 2017
- Published by Springer Nature in International Journal of Parallel Programming
- Vol. 46 (2) , 412-440
Graph Programming Interface (GPI): A Linear Algebra Programming Model for Large Scale Graph Computations
- 20 July 2012
- Published by Springer Nature in International Journal of Parallel Programming
- Vol. 41 (3) , 430-468
Efficient Stack Distance Computation for a Class of Priority Replacement Policies
- 20 July 2012
- Published by Springer Nature in International Journal of Parallel Programming
- Vol. 41 (3) , 430-468
Efficient Stack Distance Computation for a Class of Priority Replacement Policies
- 1 September 2007
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 57-65
Performance Studies of Commercial Workloads on a Multi-core System
The multi-threaded nature of many commercial applications makes them seemingly a good fit with the increasing number of available multi-core architectures. This paper presents our performance studies of a collection of commercial workloads on a multi-core system that is designed for total throughput. The selected workloads include full operational applications such as SAP-SD and IBM Trade, and popular synthetic benchmarks such as SPECjbb2005, SPEC SDET, Dbench, and Tbench. To evaluate the performance scalability and the thread-placement sensitivity, we monitor the application throughput, processor performance, and the memory subsystem of 8, 16, 24, and 32 hardware threads with (a) increasing number of cores and (b) increasing number of threads per core. We observe that these workloads scale close to linearly (with efficiencies ranging from 86% to 99%) with increasing number of cores. For scaling with hardware-threads per core, the efficiencies are between 50% and 70%. Furthermore, among other observations, our data show that the ability of hiding long latency memory operations (i.e. L2 misses) in a multi-core system enables the performance scaling.
- 1 September 2007
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 57-65
Performance Studies of Commercial Workloads on a Multi-core System
The multi-threaded nature of many commercial applications makes them seemingly a good fit with the increasing number of available multi-core architectures. This paper presents our performance studies of a collection of commercial workloads on a multi-core system that is designed for total throughput. The selected workloads include full operational applications such as SAP-SD and IBM Trade, and popular synthetic benchmarks such as SPECjbb2005, SPEC SDET, Dbench, and Tbench. To evaluate the performance scalability and the thread-placement sensitivity, we monitor the application throughput, processor performance, and the memory subsystem of 8, 16, 24, and 32 hardware threads with (a) increasing number of cores and (b) increasing number of threads per core. We observe that these workloads scale close to linearly (with efficiencies ranging from 86% to 99%) with increasing number of cores. For scaling with hardware-threads per core, the efficiencies are between 50% and 70%. Furthermore, among other observations, our data show that the ability of hiding long latency memory operations (i.e. L2 misses) in a multi-core system enables the performance scaling.
- 12 July 2004
- Published by Wiley in Software: Practice and Experience
- Vol. 34 (13) , 1257-1272
Dynamic reconfiguration of CPU and WebSphere on IBM pSeries servers
- 12 July 2004
- Published by Wiley in Software: Practice and Experience
- Vol. 34 (13) , 1257-1272
Dynamic reconfiguration of CPU and WebSphere on IBM pSeries servers