Parameter Database: Data-centric Synchronization for Scalable Machine Learning

Naman Goel; Divyakant Agrawal; Sanjay Chawla; Ahmed Khalifa Elmagarmid

doi:10.48550/arXiv.1508.00703

Parameter Database: Data-centric Synchronization for Scalable Machine Learning

Naman Goel, Divyakant Agrawal, Sanjay Chawla, Ahmed Khalifa Elmagarmid

Qatar Computing Research Institute

Hamad bin Khalifa University

Research output: Book/Report › Commissioned report › peer-review

Abstract

We propose a new data-centric synchronization framework for carrying out of machine learning (ML) tasks in a distributed environment. Our framework exploits the iterative nature of ML algorithms and relaxes the application agnostic bulk synchronization parallel (BSP) paradigm that has previously been used for distributed machine learning. Data-centric synchronization complements function-centric synchronization based on using stale updates to increase the throughput of distributed ML computations. Experiments to validate our framework suggest that we can attain substantial improvement over BSP while guaranteeing sequential correctness of ML tasks.

Original language	English
DOIs	https://doi.org/10.48550/arXiv.1508.00703
Publication status	Published - 21 Jun 2015

Access to Document

10.48550/arXiv.1508.00703

https://scholar.google.com/citations?view_op=view_citation&hl=en&user=Cb9BDJkAAAAJ&cstart=300&pagesize=100&citation_for_view=Cb9BDJkAAAAJ:7Hz3ACDFbsoC

Cite this

@book{32eb0c4606b344de9422af16520e4bfc,

title = "Parameter Database: Data-centric Synchronization for Scalable Machine Learning",

abstract = "We propose a new data-centric synchronization framework for carrying out of machine learning (ML) tasks in a distributed environment. Our framework exploits the iterative nature of ML algorithms and relaxes the application agnostic bulk synchronization parallel (BSP) paradigm that has previously been used for distributed machine learning. Data-centric synchronization complements function-centric synchronization based on using stale updates to increase the throughput of distributed ML computations. Experiments to validate our framework suggest that we can attain substantial improvement over BSP while guaranteeing sequential correctness of ML tasks.",

author = "Naman Goel and Divyakant Agrawal and Sanjay Chawla and Elmagarmid, {Ahmed Khalifa}",

year = "2015",

month = jun,

day = "21",

doi = "10.48550/arXiv.1508.00703",

language = "English",

}

TY - BOOK

T1 - Parameter Database: Data-centric Synchronization for Scalable Machine Learning

AU - Goel, Naman

AU - Agrawal, Divyakant

AU - Chawla, Sanjay

AU - Elmagarmid, Ahmed Khalifa

PY - 2015/6/21

Y1 - 2015/6/21

N2 - We propose a new data-centric synchronization framework for carrying out of machine learning (ML) tasks in a distributed environment. Our framework exploits the iterative nature of ML algorithms and relaxes the application agnostic bulk synchronization parallel (BSP) paradigm that has previously been used for distributed machine learning. Data-centric synchronization complements function-centric synchronization based on using stale updates to increase the throughput of distributed ML computations. Experiments to validate our framework suggest that we can attain substantial improvement over BSP while guaranteeing sequential correctness of ML tasks.

AB - We propose a new data-centric synchronization framework for carrying out of machine learning (ML) tasks in a distributed environment. Our framework exploits the iterative nature of ML algorithms and relaxes the application agnostic bulk synchronization parallel (BSP) paradigm that has previously been used for distributed machine learning. Data-centric synchronization complements function-centric synchronization based on using stale updates to increase the throughput of distributed ML computations. Experiments to validate our framework suggest that we can attain substantial improvement over BSP while guaranteeing sequential correctness of ML tasks.

U2 - 10.48550/arXiv.1508.00703

DO - 10.48550/arXiv.1508.00703

M3 - Commissioned report

BT - Parameter Database: Data-centric Synchronization for Scalable Machine Learning

ER -

Parameter Database: Data-centric Synchronization for Scalable Machine Learning

Abstract

Access to Document

Fingerprint

Cite this