It is increasingly common for computer users to have access to several computers on a network, and hence to be able to execute many of their tasks on any of several computers. The choice of which computers execute which tasks is commonly determined by users based on a knowledge of computer speeds for each task and the current load on each computer. A number of task scheduling systems have been developed that balance the load of the computers on the network, but such systems tend to minimize the idle time of the computers rather than minimize the idle time of the users. The paper focuses on the benefits that can be achieved when the scheduling system considers both the computer availabilities and the performance of each task on each computer. The SmartNet resource scheduling system is described and compared to two different resource allocation strategies: load balancing and user directed assignment. Results are presented where the operation of hundreds of different networks of computers running thousands of different mixes of tasks are simulated in a batch environment. These results indicate that, for the computer environments simulated, SmartNet outperforms both load balancing and user directed assignments, based on the maximum time users must wait for their tasks to finish.

/pdf/scheduling-resources-in-multi-user-heterogeneous-computing-4fepqn2z5y.pdf

Scheduling resources in multi-user, heterogeneous, computing environments with SmartNet

The author studies the performance of four mapping algorithms. The four algorithms include two naive ones: opportunistic load balancing (OLB), and limited best assignment (LBA), and two intelligent greedy algorithms: an O(nm) greedy algorithm, and an O(n/sup 2/m) greedy algorithm. All of these algorithms, except OLB, use expected run-times to assign jobs to machines. As expected run-times are rarely deterministic in modern networked and server based systems, he first uses experimentation to determine some plausible run-time distributions. Using these distributions, he next executes simulations to determine how the mapping algorithms perform. Performance comparisons show that the greedy algorithms produce schedules that, when executed, perform better than naive algorithms, even though the exact run-times are not available to the schedulers. He concludes that the use of intelligent mapping algorithms is beneficial, even when the expected time for completion of a job is not deterministic.

The relative performance of various mapping algorithms is independent of sizable variances in run-time predictions

A heterogeneous computing system provides a variety of different machines, orchestrated to perform an application whose subtasks have diverse execution requirements. The subtasks must be assigned to machines (matching) and ordered for execution (scheduling) such that the overall application execution time is minimized. A new dynamic mapping (matching and scheduling) heuristic called the hybrid remapper is presented here. The hybrid remapper is based on a centralized policy and improves a statically, obtained initial matching and scheduling by remapping to reduce the overall execution time. The remapping is non-preemptive and the execution of the hybrid remapper can be overlapped with the execution of the subtasks. During application execution, the hybrid remapper uses run-time values for the subtask completion times and machine availability times whenever possible. Therefore, the hybrid remapper bases its decisions on a mixture of run-time and expected values. The potential of the hybrid remapper to improve the performance of initial static mappings is demonstrated using simulation studies.

https://www.engr.colostate.edu/~hj/conferences/205.pdf

A dynamic matching and scheduling algorithm for heterogeneous computing systems

This paper describes and evaluates the application of three local learning algorithms-nearest-neighbor, weighted-average, and locally-weighted polynomial regression-for the prediction of run-specific resource-usage on the basis of run-time input parameters supplied to tools. A two-level knowledge base allows the learning algorithms to track short-term fluctuations in the performances of computing systems, and the use of instance editing techniques improves the scalability of the performance-modeling system. The learning algorithms assist PUNCH, a network-computing system at Purdue University, in emulating an ideal user in terms of its resource management and usage policies.

Predictive application-performance modeling in a computational grid environment

Techniques for mapping tasks to machines in heterogeneous computing systems

SmartNet is a scheduling framework for heterogeneous systems. Preliminary conservative simulation results for one of the optimization criteria, show a 1.21 improvement over Load Balancing and a 25.9 improvement over Limited Best Assignment, the two policies that evolved from homogeneous environments. SmartNet achieves these improvements through the implementation of several innovations. It recognizes and capitalizes on the inherent heterogeneity of computers in today's distributed environments; it recognizes and accounts for the underlying non-determinism of the distributed environment; it implements an original partitioning approach, making runtime prediction more accurate and useful; it effectively schedules based on all shared resource usage, including network characteristics; and it uses statistical and filtering techniques, making a greater amount of prediction information available to the scheduling engine. In this paper, the issues associated with automatically managing a heterogeneous environment are reviewed, SmartNet's architecture and implementation are described, and performance data is summarized.

D. Hensgen

Papers

SmartNet: a scheduling framework for heterogeneous computing