Browse Prior Art Database

Method of Provisioning Resources for Jobs based on Degree of Parallelism and Elasticity to Maximize Revenue in an HPC Environment

IP.com Disclosure Number: IPCOM000234543D
Publication Date: 2014-Jan-16
Document File: 2 page(s) / 41K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed system and method is to maximize revenue per unit of time in a High Performance Computing (HPC) community. The approach is to add processing nodes, without charge, to jobs with a high degree of parallelism and a low level of user-specified elasticity, so that these jobs can finish more quickly, thereby freeing resources for jobs with higher elasticity -- jobs for which users have agreed to pay for the additional running time or resources.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 51% of the total text.

Page 01 of 2

Method of Provisioning Resources for Jobs based on Degree of Parallelism and Elasticity to Maximize Revenue in an HPC Environment

The emerging cloud computing model in commercial service offerings is driving the High Performance Computing (HPC) community to offer similar levels of agility in resource provisioning in order to enable 'elasticity' of jobs. With elasticity, users in an HPC environment can request that jobs can expand in size (i.e. number of required processors), herein referred to as "spatial elasticity" or in estimated execution time, herein referred to as "temporal elasticity", on-the-fly, while the jobs are still running. Spatial elasticity of jobs allows an additional number of processors to be allocated in order to satisfy the user's resource demands, which can increase during run-time. Temporal elasticity of jobs allows the allocation of additional execution time during run-time so that jobs can run for a longer duration .

Apart from this concept of elasticity (user triggered extensions), the term 'parallelism' refers to the ability for a given job to be executed with the simultaneous use of multiple processors. Thus, the higher a given job's degree of parallelism is, the greater its ability to be simultaneously run on multiple processors. In contrast, a job that has zero parallelism is purely sequential and only executes a single task at a time on a single processor.

If systems users decide to implement (and pay for) run-time spatial and temporal elasticity, then the HPC operator must manage multiple users that submit jobs with different degrees of parallelism and elasticity requirements . A mechanism is needed

that allows the HPC operator to manage this situation while optimizing the overall revenue from users.

The solution is a novel system and method of scheduling HPC jobs in order to maximize revenue. The approach is to add processing nodes, without charge, to jobs

with a high degree of parallelism (i.e., jobs that can benefit from additional processors ) and a low level of user-specified elasticity, so that these jobs can finish more quickly. Thus, resources...