19-12-2012, 05:39 PM
Libra: An Economy-Driven Cluster Scheduler Software Requirements Specification
[attachment=44467]
Introduction
Purpose
This SRS describes the requirements and specifications of Libra, an Economy-Driven Cluster Scheduler. It explains the functional features of the scheduler, along with interface details, design constraints and related considerations such as performance characteristics. The SRS is intended for users and owners of high-performance clusters, cluster management software, job schedulers and grid resources.
Scope
The Libra Scheduler is intended to work as an add-on to the queuing, scheduling and resource managing module of the open source Sun Grid Engine cluster management system, previously known as Codine. The scheduler will offer market-based economy driven service for managing batch jobs on clusters by scheduling CPU time according to user utility rather than system performance considerations. Hence, the main objective of Libra is to provide Quality of Service (QoS) computational economy in cluster computing.
As part of the project, the Libra Scheduler will only manage sequential and embarrassingly parallel batch jobs to be run on a homogenous Linux cluster, although its functionality can be adapted to schedule resources other than CPU time, even on a heterogeneous cluster.
To provide QoS to users, there will be no mechanism for users to interact with each other, and bargain on the use of resources according to their considerations, as is provided in a grid-computing environment by projects like Nimrod/G [1].
Definitions, Acronyms and Abbreviations
Cluster/Cluster Computing
Clustering involves connecting two or more computers together to take advantage of combined computational power and resources. Hence, a cluster works as an integrated collection of resources that can provide a single system image spanning all its nodes. Clustering is a popular strategy for processing applications because it transparently spreads the processing of different jobs throughout the cluster, and are used for high-performance applications such as AI expert systems, nuclear simulations, and scientific calculations.
Grid Computing [3]
A grid is an infrastructure that couples resources such as computers (PCs, workstations, clusters and so on), software (for special purpose applications) and devices (printers, scanners) across the Internet and presents them as a unified integrated single resource that can be widely used. Grid computing refers to using a grid for processing jobs and processes.
Bid-based Proportional Resource-Sharing Model [4]
This is an economic model that specifies one way of allocating (cluster) resources between competing users based on what their criteria is, such as deadline and budget. In this model, the amount of resource allocated to users is proportional to the values of their bids. Essentially, in a bid-based system, the users are allocated tokens or credits, which they can use for having access to resources. The value of each credit depends on the resource demand and cost at the time of usage. We can modify this approach for our cluster, by evaluating this credit/priority according to the deadline and price constraints of users, and considering the system load and resources use as well. This priority is then used by a scheduling policy to allocate resource proportionally to the competing jobs.
Stride Scheduling [5]
This is a scheduling algorithm to enforce resource allocations proportional to a user’s priority or share. A pool of tickets represents resource rights to a shared resource. A stride is inversely proportional to the number of tickets allocated to a job, hence representing the interval (in terms of CPU time) between two successive scheduling of a job. Jobs with more tickets have shorter strides and hence are scheduled more frequently and obtain a greater fraction of the resource. The algorithm is deterministic and quite accurately schedules jobs in exact proportions based on their ticket allocations.
[attachment=44467]
Introduction
Purpose
This SRS describes the requirements and specifications of Libra, an Economy-Driven Cluster Scheduler. It explains the functional features of the scheduler, along with interface details, design constraints and related considerations such as performance characteristics. The SRS is intended for users and owners of high-performance clusters, cluster management software, job schedulers and grid resources.
Scope
The Libra Scheduler is intended to work as an add-on to the queuing, scheduling and resource managing module of the open source Sun Grid Engine cluster management system, previously known as Codine. The scheduler will offer market-based economy driven service for managing batch jobs on clusters by scheduling CPU time according to user utility rather than system performance considerations. Hence, the main objective of Libra is to provide Quality of Service (QoS) computational economy in cluster computing.
As part of the project, the Libra Scheduler will only manage sequential and embarrassingly parallel batch jobs to be run on a homogenous Linux cluster, although its functionality can be adapted to schedule resources other than CPU time, even on a heterogeneous cluster.
To provide QoS to users, there will be no mechanism for users to interact with each other, and bargain on the use of resources according to their considerations, as is provided in a grid-computing environment by projects like Nimrod/G [1].
Definitions, Acronyms and Abbreviations
Cluster/Cluster Computing
Clustering involves connecting two or more computers together to take advantage of combined computational power and resources. Hence, a cluster works as an integrated collection of resources that can provide a single system image spanning all its nodes. Clustering is a popular strategy for processing applications because it transparently spreads the processing of different jobs throughout the cluster, and are used for high-performance applications such as AI expert systems, nuclear simulations, and scientific calculations.
Grid Computing [3]
A grid is an infrastructure that couples resources such as computers (PCs, workstations, clusters and so on), software (for special purpose applications) and devices (printers, scanners) across the Internet and presents them as a unified integrated single resource that can be widely used. Grid computing refers to using a grid for processing jobs and processes.
Bid-based Proportional Resource-Sharing Model [4]
This is an economic model that specifies one way of allocating (cluster) resources between competing users based on what their criteria is, such as deadline and budget. In this model, the amount of resource allocated to users is proportional to the values of their bids. Essentially, in a bid-based system, the users are allocated tokens or credits, which they can use for having access to resources. The value of each credit depends on the resource demand and cost at the time of usage. We can modify this approach for our cluster, by evaluating this credit/priority according to the deadline and price constraints of users, and considering the system load and resources use as well. This priority is then used by a scheduling policy to allocate resource proportionally to the competing jobs.
Stride Scheduling [5]
This is a scheduling algorithm to enforce resource allocations proportional to a user’s priority or share. A pool of tickets represents resource rights to a shared resource. A stride is inversely proportional to the number of tickets allocated to a job, hence representing the interval (in terms of CPU time) between two successive scheduling of a job. Jobs with more tickets have shorter strides and hence are scheduled more frequently and obtain a greater fraction of the resource. The algorithm is deterministic and quite accurately schedules jobs in exact proportions based on their ticket allocations.