18-07-2012, 03:47 PM
ROBUST LOAD DELEGATION IN SERVICE GRID ENVIRONMENT
[attachment=27985]
INTRODUCTION
Today, the use of Grid Computing is not anymore limited to HPC/HTC-centric communities such as High Energy Physics, Astronomy, or Climate Research, which have a certain tradition of using such infrastructures. Other sciences—e.g., Financial Services, Construction Engineering, and even arts and humanities—also start to adopt Grid Computing as a tool for e-Science, and show an ever-increasing demand for computing power and storage space. While well-established approaches such as the EGEE environment have relied on centralized middleware infrastructures for whole e-Science communities, other— mostly emerging—efforts, such as the D-Grid Initiative, I have chosen a Service Grid approach with smaller, more community-tailored Grids. In the latter case, however, a strong demand for enabling collaboration and cooperation on the infrastructure layer between the different communities and Grids can be observed.
Modules:
Information Policies
LRMS (local resource management systems)
Adaptive Decision Making
Load Delegation.
Information Policies
This policy becomes relevant if more than one exchange partner is available in the Grid. Thus, there exists more than one possibility to delegate a job to a remote Grid participant. For such scenarios, the location policy determines as a first step the sorted subset of possible delegation targets
Transfer Policy
After the location policy has been applied, the transfer policy specifies whether a job should be delegated to a certain partner or not. For this purpose, the policy is applied separately on each partner in a predetermined order. Every time the transfer policy is consulted, it decides whether the job should be executed locally or delegated to the considered partner.
LRMS (local resource management systems)
The LRMS layer consists of a waiting queue and a scheduler. The waiting queue stores all locally submitted jobs while the scheduler executes a specific scheduling strategy in order to assign jobs from the waiting queue onto the available local resources. On MPP system layer, this approach allows the realization of priorities for jobs of different user groups. Usually, the scheduling strategies are formulated by the system provider to fulfill the users’ needs. Although many special-purpose algorithms exist that are tailored for certain MPP system owner priorities, we use the basic and simple First-Come-First-Serve (FCFS) algorithm as an example on LRMS. This heuristic starts the first job of the waiting queue whenever enough idle resources are available.
Adaptive Decision Making
The current state of the system is crucial when deciding on whether to accept or decline foreign workload, e.g., allowing for additional remote jobs, if the local system is already highly loaded, seems to be inappropriate.
Dependent on the current system state, the Fuzzy decision maker has to decide whether to accept an offered job or not. Thus, we represent the acceptance of a job by an output value of 1 and the corresponding refusal of a job by 1.
Existing System:
Early approaches favor a hierarchical scheduling structure, where a central scheduler instance—often called Metascheduler, Grid Scheduler, or Broker—delegates submitted jobs to subordinated partner sites. Such a centralized model implies full knowledge of the Grid sites’ state and exclusive control over the LRMS to facilitate efficient job scheduling. The afore described approach is contrasted by a decentralized structure, in which local sites can act as autonomous peers and share jobs in an equitable fashion. Thus, the process of job interchange is deferred to the competence of each local site scheduler and puts special emphasis on the decision-making process of accepting remote and dispensing local jobs. With respect to the basic parameters of modern
[attachment=27985]
INTRODUCTION
Today, the use of Grid Computing is not anymore limited to HPC/HTC-centric communities such as High Energy Physics, Astronomy, or Climate Research, which have a certain tradition of using such infrastructures. Other sciences—e.g., Financial Services, Construction Engineering, and even arts and humanities—also start to adopt Grid Computing as a tool for e-Science, and show an ever-increasing demand for computing power and storage space. While well-established approaches such as the EGEE environment have relied on centralized middleware infrastructures for whole e-Science communities, other— mostly emerging—efforts, such as the D-Grid Initiative, I have chosen a Service Grid approach with smaller, more community-tailored Grids. In the latter case, however, a strong demand for enabling collaboration and cooperation on the infrastructure layer between the different communities and Grids can be observed.
Modules:
Information Policies
LRMS (local resource management systems)
Adaptive Decision Making
Load Delegation.
Information Policies
This policy becomes relevant if more than one exchange partner is available in the Grid. Thus, there exists more than one possibility to delegate a job to a remote Grid participant. For such scenarios, the location policy determines as a first step the sorted subset of possible delegation targets
Transfer Policy
After the location policy has been applied, the transfer policy specifies whether a job should be delegated to a certain partner or not. For this purpose, the policy is applied separately on each partner in a predetermined order. Every time the transfer policy is consulted, it decides whether the job should be executed locally or delegated to the considered partner.
LRMS (local resource management systems)
The LRMS layer consists of a waiting queue and a scheduler. The waiting queue stores all locally submitted jobs while the scheduler executes a specific scheduling strategy in order to assign jobs from the waiting queue onto the available local resources. On MPP system layer, this approach allows the realization of priorities for jobs of different user groups. Usually, the scheduling strategies are formulated by the system provider to fulfill the users’ needs. Although many special-purpose algorithms exist that are tailored for certain MPP system owner priorities, we use the basic and simple First-Come-First-Serve (FCFS) algorithm as an example on LRMS. This heuristic starts the first job of the waiting queue whenever enough idle resources are available.
Adaptive Decision Making
The current state of the system is crucial when deciding on whether to accept or decline foreign workload, e.g., allowing for additional remote jobs, if the local system is already highly loaded, seems to be inappropriate.
Dependent on the current system state, the Fuzzy decision maker has to decide whether to accept an offered job or not. Thus, we represent the acceptance of a job by an output value of 1 and the corresponding refusal of a job by 1.
Existing System:
Early approaches favor a hierarchical scheduling structure, where a central scheduler instance—often called Metascheduler, Grid Scheduler, or Broker—delegates submitted jobs to subordinated partner sites. Such a centralized model implies full knowledge of the Grid sites’ state and exclusive control over the LRMS to facilitate efficient job scheduling. The afore described approach is contrasted by a decentralized structure, in which local sites can act as autonomous peers and share jobs in an equitable fashion. Thus, the process of job interchange is deferred to the competence of each local site scheduler and puts special emphasis on the decision-making process of accepting remote and dispensing local jobs. With respect to the basic parameters of modern