Mark E. Schaefer

Auction allocation of computing resources

Standard methods for allocating computing resources normally employ schedulers and either queue or priority schemes. Alternative methods utilizing marketlike processes are being investigated, with direct applicability to evolving distributed systems. In this article, we present results of simulations of an auction allocation is which computing tasks are provided sufficient intelligence to acquire resources by offering, bidding, and exchanging them for funds.

Public and private optimization at a service facility with approximate information on congestion

Abstract An M / GI /1 queueing model is considered, where the arrival rate to the facility is a continuous variable which depends, in the steady state, upon the average congestion at the facility. The population of customers arriving to the facility is partitioned into several classes dependent on the ratio of the value of time to the reward due to service but are served according to first-in-first-out rule. It is shown that under the privately optimal behavior of the individuals the facility will be dominated by the class with the highest net reward per value of time. The publicly optimal policy which maximizes the net reward due to service, after costs of waiting are deducted, is shown either to admit only a single class of customers to the facility, thus discriminating against the other classes or to be indifferent to the mix of classes. The class chosen for admission may not be the class which would have privately dominated the facility. When the expected delay experienced at the facility is fixed, a policy of tolls and rebates for the customers is obtained that will assure equal access to the facility for all customers irrespective of their classes. It is shown that the publicly optimal policy, under the condition of fixed aggregate arrival rate to the facility, is shown to be deversified. The optimal arrival rates desired by a single class are derived for two cases. When the proportions of arrivals from the classes are fixed, the aggregate arrival rate desired by a class is shown to be not greater than the equilibrium rate for the individuals from that class. Alternatively, when the aggregate arrival rate is fixed, conditions are obtained under which a class will prefer usage of the facility by several classes to its own domination.

Modeling The Cost Of Resource Allocation In Distributed Control

Modifying our previously developed simulation model [FRA89], we study in this paper the costs associated with distributed allocation of computing resources in a multitasking environment. Using funds endowed upon arrival, computing tasks compete for necessary resources through sealed-bid auctions to improve their processing schedules. The costs and times dedicated to auctioning are compared to the costs and times allowed for task processing. Measuring computing resources in terms of processing rates allows the task management, in the form of an auction, algorithm, to have its requirements specified in the same way as the requirements for the simulated mission processing. Machine capacity is computed for and assigned to each completing task. Data are then compiled by segmented capacity classes. A unifying theme of past and current research is the efficiency of auctioning to allocate reconfigurable computing resources in a variable capacity machine. We observed that at optimal rates of occurrence of capacity classes which minimize the total costs per successful completion, congestion was resolved through auctions generating endogenously implied prices which substantially exceeded the exogenously imposed price.

The simulation of a distributed control model for resource allocation and the implied pricing

The allocation of computing resources and the scheduling of tasks in a multitasking environment are simulated using a distributed control model. The tasks compete for computing resources in a decentralized manner through sealed bid auctions to improve their schedules, rather than having resources centrally administered by a host controller. Funds used for bidding are endowed to the tasks upon arrival at the computing system. The effects on completion times of three endowment strategies and two machine sizes are analyzed using a range of system capacities. Within each capacity class, an apparent cost, derived from the run parameters, is contrasted with an implied price generated by the auction process. Performance is examined in terms of congestion at various capacities. At optimal (lowest cost per successful completion) rates of occurrence of these capacity classes, an implied price arises that exceeds the “free access” price. This internally generated price appears to ration resources and time, thus discouraging congestion. Implementing such a distributed control algorithm suggests that determining a price schedule for allocating computing resources can be moved “to the left” in the system life cycle.

Optimal acceptance of job orders

Consider a production centre receiving orders for several products. Assuming the first two moments of the processing times of the products as known and given the profit per unit of each product, the model provides a basis for selecting the optimal mix of job orders to be accepted. In particular, the main feature of the optimal solution is that it will in most cases call for accepting job orders from a wide spectrum of classes

The simulation of decentralized control: a hostless resource allocation model

In this article, an alternative approach to control for computing systems, with possible distributed, parallel, or multilprocess application, is proposed and evaluated through simulation. Functions normally handled by centralized controllers, schedulers, arbiters and priority schemes are accomplished through a decentral ized model of control. Resource allocation, one important control function, is resolved within a Challenge Ring (CR) in which individual computing tasks independently (or without a host, hence their interaction is called hostless) exercise algorithms to gain access to computing resources. Simulated system performance is monitored by analyzing individual task processing times, total system times, resource availability, resource utilization, and system efficiency. Our preliminary experimental results indicate that such decentralized (or hostless) models can be superior to some standard centralized (or hosted) versions. Moreover, tasks in CR networks that interact through cooperative strategies in some cases exhibit better performance. Our overall results encourage the further exploration of decentralized control models which could be useful in the continuing pursuit of alternative machine constructs (e.g. non-von Neumann architectures) and new distributed operational schemes (e.g. hostless network operating systems).

Functionality in the reusability of software

Reusable mathematical and statistical software has been available for some time providing substantial programming productivity to scientists and engineers. With the development of Ada and extensions of its package concept to more general categories of information systems, libraries of reusable code and cataloging schemes are anticipated from regular practices to yield the same kinds of benefits to these systems. One obstacle has been the lack of abstractions for identifying and developing candidate reusable software, promoting its later classification, recovery and modification. This paper advances a method to discover common functions across different systems to enhance reusability.