An EasyGrid portal for scheduling system‐aware applications on computational Grids

Abstract

SUMMARYOne of the objectives of computational Grids is to offer applications the collective computational power of distributed but typically shared heterogeneous resources. Unfortunately, efficiently harnessing the performance potential of such systems (i.e. how and where applications should execute on the Grid) is a challenging endeavor due principally to the very distributed, shared and heterogeneous nature of the resources involved. A crucial step towards solving this problem is the need to identify both an appropriate scheduling model and scheduling algorithm(s). This paper presents a tool to aid the design and evaluation of scheduling policies suitable for efficient execution of system-aware parallel applications on computational Grids.

Distributed and dynamic self‐scheduling of parallel MPI Grid applications

“…Since the cost of scheduling an application at compile time is generally neglected, more sophisticated heuristics can be employed [3]. By taking advantage of the static scheduler's ability to analyze the whole application and the dynamic scheduler's access to accurate runtime system information, more informed scheduling decisions can be made quickly.…”

“…computing power, communication costs) be captured by an architectural model. In a DAG G = (V , E, ε, ω): the set of vertices, V , represent tasks; E is the precedence relation among the tasks; ε(v) is the amount of work associated to task v ∈ V ; and ω(u, v) is the weight associated to the edge (u, v) ∈ E, representing the amount of data transmitted from task u to v. Given the set of Grid machines the user has access to and their availability, the EasyGrid scheduling portal executes an MPI modelling program with the user's Grid credentials (the Grid community has adopted a public key infrastructure for authentication on Grid resources) to estimate the likely performance available to the user's application [3]. This information forms the architectural model upon which, together with a DAG representation of the application, the portal employs a selection of scheduling heuristics to find a good initial schedule [3].…”

“…Finally, a host manager (HM) on each machine is responsible for the scheduling, creation and execution of the application processes associated with that host. For usability, the task of transforming a cluster-based application into a Smart G-App equivalent and initiating its execution is carried out automatically by the EasyGrid scheduling portal [3]. The portal is capable of: choosing the appropriate AMS scheduling and fault-tolerance policies for the application; determining an initial process allocation; compiling the Smart G-App application; managing the user's Grid proxy; creating the MPI environment (including transferring files); and providing fault tolerance for the GM process (should this process fail, the portal must intervene to restart the application).…”

“…The AMS is embedded automatically into a user's parallel MPI application without modification of the original code by a scheduling portal [3] which, as with almost all management systems, also leverages the core middleware functionalities offered by the Globus Toolkit [4] to facilitate the use of Grid computing. The implementation of an AMS and the viability of transforming legacy clusterbased MPI applications into system-aware applications capable of executing in Grid environments was discussed in [5].…”

Efficient hierarchical self-scheduling for MPI applications executing in computational Grids

“…The AMS is embedded automatically into a user's parallel MPI application without modifications to the original code by a scheduling Portal which also acts as a simple RMS [3]. The implementation of an AMS and the viability of transforming legacy cluster based MPI applications into system-aware ones capable of executing in grid environments was discussed in [11].…”