Home About Software Documentation Support Outreach Ecosystem Blog Dev Awards Team & Sponsors

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

about:functionality:schedcapab [ - 2015/03/09, 12:58 - ] (current)
Line 1: Line 1:
 +
 +====== A. Scheduling Capabilities ======
 +
 +===== A.1. Dynamic Scheduling =====
 +
 +**Description:​**
 +
 +    * Grid resources are highly dynamic, both in load and availability. Dynamic scheduling matches job requirements to the static and dynamic load attributes of the resources available in each scheduling cycle.
 +
 +**Support in Last Release:**
 +
 +    * GW supports dynamic scheduling. Each job provides the requirements that must be met by the target resources (discovery process) and criteria to rank the matched resources (selection process). The requirement and rank expressions are based on common resource attributes, independent from the information service.
 +    * Different Information management drivers are provided to access the more common information services.
 +    * GW modular architecture allows to interface with any information service.
 +
 +
 +===== A.2. Opportunistic Migration =====
 +
 +
 +**Description:​**
 +
 +    * Opportunistic migration means automatic resubmission of running jobs to better resources. The scheduler evaluates the performance benefits that can be obtained due to the migration of the job.
 +
 +**Support in Last Release:**
 +
 +    * Each scheduling cycle, GW evaluates the benefits of migrating submitted jobs to new available resources (recently added or freed) by comparing rank values.
 +
 +===== A.3. Performance Slowdown Detection =====
 +
 +
 +**Description:​**
 +
 +    * The scheduler requests a migration when a intolerable performance loss for a job (performance contract violation) is detected. Grid users access resources that are being exploited by other grid users, as well as by internal users. Some local job managers do not guarantee exclusive access to compute resources, initially idle hosts may become saturated, and vice versa. On the other hand, in dedicated batch systems, the saturation of the resource may increase the queue wait time to an unacceptable value.
 +
 +**Support in Last Release:**
 +
 +    * GW counts the suspension time in remote batch systems and requests a migration when it exceeds a given threshold.
 +    * Jobs are submitted together with a light-weight self monitoring system. The job will be migrated when it is not receiving as much CPU as you expected.
 +
 +===== A.4. Self-adaptive Applications =====
 +
 +
 +**Description:​**
 +
 +    * The emerging of Grid technology has led to a new generation of applications that relies on its ability to adapt its execution to dynamic Grid conditions. These new applications must be able to seek out computational resources that fit their needs as their execution evolves. For instance, adaptive-mesh refinement numerical methods systematically refine the computational mesh in those areas where a higher resolution is needed. In this sense, the amount of RAM memory needed to store the computational mesh is not known beforehand. The application must adapt itself to its new requirements,​ migrating to a resource that provides an enough amount of RAM memory.
 +
 +**Support in Last Release:**
 +
 +    * An application can take decisions about resource selection as its execution evolves by modifying its requirement and rank expressions and requesting a migration.
 +
 +===== A.5. Checkpointing Support =====
 +
 +
 +**Description:​**
 +
 +    * Checkpointing is required to support migration on-request, opportunistic migration and fault tolerance.
 +
 +**Support in Last Release:**
 +
 +    * User-level checkpointing or restarted architecture-independent files managed by the programmer must be implemented because system-level checkpointing is not possible among heterogeneous resources. Migration is implemented by restarting the job on the new candidate host. If the checkpointing files are not provided the job should be restarted from the beginning.
 +    * Check-points periodically retrieve to the client machine or a checkpoint server.
 +
 +===== A.6. Scheduling Policies =====
 +
 +
 +**Description:​**
 +
 +    * Scheduling policies define the balance of workload among the available resources.
 +
 +**Support in Last Release:**
 +
 +    * GW provides state-of-the-art scheduling policies, comprising job prioritization policies (fixed priority, urgency, share, deadline and waiting-time) and resource prioritization policies (fixed priority, usage, failure and rank).
 +
 +===== A.7. Scheduling Reporting and Accounting =====
 +
 +
 +**Description:​**
 +
 +    * Scheduling reporting and accounting facilities provide information about overall performance and help troubleshoot configuration problems.
 +
 +**Support in Last Release:**
 +
 +    * GW generates scheduling and job logging.
 +    * Accounting of resource consumption (average execution and transfer times, failures...) and job type execution.
 +
 +===== A.8. Job Dependency =====
 +
 +
 +**Description:​**
 +
 +    * Job dependency capability allows the execution of a submitted job depending on the completion of other jobs submitted in the grid.
 +
 +**Support in Last Release:**
 +
 +    * GridWay allows job submission to be dependent on the completion of other jobs. This new funcionality provides support for the execution of workflows. ​
 +
 +===== A.9. Single and Multiple-user Support =====
 +
 +
 +**Description:​**
 +
 +    * Single-user meta-schedulers provide job execution management and resource brokering for one user (user-level schedulers for partner grids). On the other hand, multiple-user meta-schedulers provide support for multiple users in each scheduling instance (centralized schedulers for enterprise grids or VO-level schedulers for partner grid). ​
 +
 +**Support in Last Release:**
 +
 +    * Single-user mode support. GridWay can be installed by the end-user in his client host.
 +    * Multiple-user mode support. This way the installation and configuration of GridWay can be performed by the system manager and the users could submit, control and monitor their jobs from a front-end or from client hosts (that do not require a GridWay and Globus installation). The administrator can define scheduling policies allowing prioritization of jobs and users. For example, fair-share scheduling among users. ​
  
Admin · Log In