A guide to simulation (2nd ed.)
A guide to simulation (2nd ed.)
Adaptive performance prediction for distributed data-intensive applications
SC '99 Proceedings of the 1999 ACM/IEEE conference on Supercomputing
Virtual Computing: Concept, Design, and Evaluation
Virtual Computing: Concept, Design, and Evaluation
Simgrid: A Toolkit for the Simulation of Application Scheduling
CCGRID '01 Proceedings of the 1st International Symposium on Cluster Computing and the Grid
Hi-index | 0.00 |
Grid computing-- the assemblage of heterogeneous distributed clusters of computers viewed as a single virtual machine-- promises to serve as the next major paradigm in distributed computing. Since Grids are assemblages of (usually) autonomous systems (autonomous clusters, supercomputers, or even single workstations) scheduling can become a complex affair which must take into consideration not just the requirements (and scheduling decisions) made at the point of the job's origin, but also the scheduling requirements (and decisions) made at remote points on the fabric, and in particular scheduling decisions made by a remote autonomous system onto which the local job has been scheduled. The current existing scheduling models range from static, where each of the programs is assigned once to a processor before execution of the program commences, to dynamic, where a program may be reassigned to different processors, or a hybrid approach, which combines characteristics of both techniques [1,4,5].To address this issue, we have developed a JAVA based discrete event Grid simulator toolkit called HuskySim. The HuskySim toolkit provides core functionalities (e.g., compute objects, network objects, and scheduling objects) that can be used to simulate a distributed computing environment. Furthermore, it can be used to predict the performance of various classes of Grid scheduling algorithms including: Static scheduling algorithms, Dynamic scheduling, Adaptive Scheduling.In our design, we adopted an object-oriented design, which allows an easy mapping and integration of simulation objects into the simulation program. This approach simplifies the simulation of multitasking, and distributed data processing model. Our model of multitasking processing is based on an interrupt driven mechanism.As shown in Figure 1, the simulator works by relaying messages between the core engine and the simulation modules through the message handling sub-system. Once the architecture, the load distribution, and the scheduling algorithms are defined, the object registration subsystem sends a NEW OBJECT REQUEST MESSAGE to the object class libraries and builds a skeleton for the requested simulation experiment.Workloads traces can be generated using probabilistic models. The currently supported distributions are: Uniform, Poisson, Exponential, Normal, Erlang, and Power Tailed. It is also possible to use real world load traces. Moreover, we augmented the Simulator with a statistical module. Using the statistical module provided with the HuskySim, the core simulation engine can send messages to perform various type of analysis on the performance data including: variance reduction, regression, time series analysis, clustering, and data mining.In order to quantify the system performance, the simulator provides various performance metrics including: CPU utilization, disk utilization, application turnaround time, latency, make span, host to host bandwidth, jammed bandwidth, and TCP/IP traffic data. These measurements are handled through the measurement sub-system.Furthermore, the HuskySim can be used to simulate the classes of algorithmic and parametric adaptive Grid schedulers. In which, the scheduling algorithm may not be fixed in advance. Simply, the scheduling algorithm is selected at run time based on the current workload on the Grid fabric in order to operate at near optimal level.