Citation Context

...nd works at the process level. However, the algorithms are themselves independent of the application. Recent work however has suggested extending the GrADS scheduling concept to workflow applications =-=[144]-=-. However, the treatment of data still remains the same. Ranganathan and Foster [80] propose a decoupled scheduling architecture for data intensive applications which consists of 3 components: the Ext...

Citation Context

...0.085/hour to $2.10/hour for Amazon EC2). We define a VM type in terms of the estimated processing time for each service unit, the cost per unit time (currently it is 1 hour for cloud providers) and the estimated instance acquisition lag. We assume there is no performance inference among VMs. VMv = {[ ]v , cv , lagv} Performance is a vector, in which each element is the estimated processing time for tasks from service unit Si. Such performance estimation on different types of VMs can be achieved by using existing performance estimation techniques [16] (e.g. analytical modeling [33], empirical [34] and historical data [35][36]). Furthermore, we know the cost of running task on VMv is cv. 3.2.3 The auto-scaling problem Definition 3 (Workload). We assume the application owner does not know the incoming requests in advance. Therefore, the workload is defined as all the jobs that have been submitted into the application. They are either waiting to be processed or partially processed. At some time t, we use the tasks waiting at every service unit to represent the workload W. Wt = ∑ ∑ Definition 4 (Scheduling Plan). The auto-scaling mechanism needs to balance two decisions. The first decision...

Citation Context

...encies among computing tasks in scientific workflows is to use a Directed Acyclic Graph (DAG). Prominent grid projects using DAGs are the Pegasus system from GriPhyn project [15], Taverna [27], GrADS =-=[12]-=-, GridFlow [8] and GridBus [7]. Scripting languages are another approach, such as Matrix from SDSC [37], GridAnt [5] and JPhython in XCat [24], and BPEL4WS (Business process Execution Language for Web...

Citation Context

...ing USC, on his sabbatical leave from the University of Hong Kong (HKU). Kwok’s research at USC was also supported by a research grant from the Research Grants Council of the HKSAR and the 2003-2004 Outstanding Young Researcher Award given by HKU. 1. Introduction Grid computing will eventually become commodities that are transported among different Grid sites (which can be large supercomputing centers or just small clusters) and executed by the sites in a transparent manner [4, 6]. Largescale Grid computing demands a judicious job management system to address the security and privacy concerns [5, 9, 10, 14]. A job dispatched to a remote site may fail due to experiencing some infections or malicious attacks there. Specifically, a practical job scheduler must be security-driven in that it must consider the risk involved in dispatching jobs to remote sites. Unfortunately, existing Grid scheduling algorithms largely ignore this issue, making their applicability in a realistic environment rather doubtful. Grids are most formed with resources owned by many organizations and thus are not dedicated for certain users. As such, jobs that are dispatched to a remote site can possibly experience some securit...

Citation Context

...tion to default RMA scheduling algorithms, researchers have proposed pluggable scheduling modules to extend the scheduling capabilities of the resource manager. These scheduling systems include GrADS =-=[7]-=-, Pegasus [8], Taverna [9], and ICENI [10]. Among these systems only Condor-G and ICENI provide an advance reservation mechanism in their schedulers but both of them do not consider backfilling while ...

Citation Context

...trol or data dependencies. As a result, a number of Grid workflow management systems with scheduling algorithms have been developed by several projects such as Condor DAGMan [22], Askalon [11], GrADS =-=[8]-=-, ICENI [16], APST [3], and Pegasus [4][9]. They facilitate the execution of workflow applications and minimize their execution time on Grids. However, for imposing workflow paradigm on utility Grids,...

Citation Context

...m (also referred as allocation, mapping or partitioning) has been studied for many years with various models. It is an essential research topic in the field of parallel computing [11], grid computing =-=[17]-=-, and more recently, in the field of distributed database [18]. In [19], Stone proposed the optimal task assignment algorithm on two-processor problem by using Max-flow/Min-cut theorem. This work laid...