Академический Документы
Профессиональный Документы
Культура Документы
47
3. Schedule resource based on PS systems receiving an allocation xi .We make the following
assumptions about brokers' valuations:
In Nimrod-G, a user specifies QoS parameters, such as Assumption 1:For all i ∈ {1,...N }
deadline and budget to a broker. Then, the broker
schedule user’s tasks to resources with different vi ( xi ) is continuously differentiable
allocations system, which does not use a proportional vi′ ( xi ) > 0 ; ∀xi ∈ (0,1)
share schedule system. GridSim[3] adopted Post Price The first assumption captures the fact that broker's
model as its economic model, in time optimization performance or marginal valuation of performance should
scheduling algorithm and cost optimization scheduling not change dramatically given a slight change in
algorithm, it just simply assigns one or many jobs to the allocation. The second assumption is intuitive as a
one resource only, the resource occupy type are broker's valuation should increase with allocation. Each
exclusively occupy it , therefore it only considers about broker's utility is the difference between the valuation and
the scenario of job in sequence queue. It does not cost of its allocation:
considers the problem of job competition, however, in the U i ( s ) = vi ( xi ( s)) − ci ( s , x)
real Grid environments, the occupy type of resource are
mostly share type. Moreover, users from different VOs Substituting from Equations (4) and (5), we have:
may competing to use the same resource at the same time. si (6)
U i ( s ) = U i ( si , s− i ) = vi ( ) − si
In fact, maximizing utilization and ensuring fairness si + s− i + ε
among users are two major issues in resource allocation. where s− i = ∑ N −1 s j − si is the sum of the bids of all
A system that uses a proportional share algorithm [7, 8], j =1
calculates a user share of a resource based on a user agents excluding the i-th agent and sN = ε is a bid made
weight in relation to the total weight of all users in the by an agent representing the resource. By bidding ε , the
system. However, a normal proportional share system resource has a way of declaring a reservation value for its
does not check the validity of each user weight. This will resource and prevents the possibility of agents colluding
lead to incorrect priority when one user gives a low- to purchase the resource for an arbitrarily small amount of
priority task the same weight as high-priority tasks of money. The first order necessary condition for a
other users. In this paper, the proportional share system maximizing interior solution is:
based on the user’s bidding price are proposed, therefore,
users with a tight deadline will bid higher in order to gain U i′ ( si , s− i ) = vi′( xi ( s )) xi′ ( s ) − 1
more resource shares. si si
= vi′ ( ) −1 = 0
3.1 Allocation Mechanism si + s− i + ε ( si + s− i + ε ) 2
The Grid allocation problem can be formulated as Can be writing in to:
follows:
si ( s + s + ε )2
At first, N brokers of users competing for a resource vi′ ( ) − i −i =0
with fixed finite capacity. The resource is allocated using si + s− i + ε s− i + ε
market mechanism, where the resource schedule depends The left hand side of the above equation is a decreasing
on the relative signals or bids sent by the brokers. We function of si as vi′ (.) is decreasing in its argument,
assume that each broker submits a signal si to the
si / si + s− i + ε is an increasing function of si and the
resource. In our proportional share mode, we want our
allocations to be proportionally fair by bidding price. This second term has si only in the numerator. So, an interior
can be achieved with the following allocation rule: solution exists if and only if the left hand side is positive
s (4) when si = 0. A broker will participate in the bid (i.e.,
xi ( s ) = i
∑j i
s submits a nonzero bid), if and only if:
We note that this rule satisfies the proportionally fair vi′ (0) > s− i + ε
criterion for allocation in the systems proposed for
computational resources. In terms of cost of computation, In the application, each user has a broker that is
we note that it takes O(N) operations to perform the responsible in monitoring the progress of application and
allocation presented in equation (4), which is the minimal managing on how much to bid on each resource. A user
cost for making variable allocations to N brokers. If the can be part of many VO domains [15]. We consider a
broker has no extra profit, the cost for each broker is scenario where brokers are generated at some subset (user
request) of nodes in Grid with a sequence of jobs to
ci ( s, x ) = si (5) complete. The jobs require access to resources available
3.2 Broker Utility at various nodes throughout the Grid. Based on some
We assume that each broker has a valuation vi ( xi ) for budget constraints, the broker attempts to purchase
resources throughout the Grid to complete its set of jobs
48
according to a given performance measure or utility ∂L qik − qik s−ki k 2
by the i-th broker for its k-th task is min ∑ tik s.t. ∑e k
i ≤ Ei
k =1 k =1
sk
xik = Cik ( k i k ) 3.5 Deadline and Cost Optimization Algorithm
si + s−i We can define the problem as the agent balance its
Then, the time taken to complete that job will be performance as measured by the time taken to complete
qk (s k + s k ) its jobs and the cost of obtaining service. We can consider
tik = i i k k − i it as the following criterion:
Ci si Ki Ki
The expense to the broker is the bid times the duration min ∑ α ik eik + (1 − α ik )∑ tik
of service, which yields k =1 k =1
49
based on proportional system that schedules a user
application in resource allocation systems.
In the future, we plan to design the system more details.
Moreover, we are planning to incorporate resource
reservation and different economic models into resource
allocation. This allows us to contrast the efficiency of
different economic models. In addition, we are
considering credit evaluation and job preemption and job
migration policy that tries to achieve the integrity of
market transactions.
6. References
[1] D. P. Anderson, J. Cobb, E. Korpela, M. Lebofsky,
and D. Werthimer. SETI@home: an experiment in
public resource computing. Commu.. ACM,
45(11):56–61, 2002.
[2] I. Foster and C. Kesselman (editors), The Grid:
Blueprint for a Future Computing Infrastructure,
Figure 1 Average time and average cost in different Morgan Kaufmann Publishers, USA, 1999. pp. 123–
number resource and different strategy 135.
It can be observed that Deadline and Cost [3] R. Buyya, D. Abramson, and J. Giddy, A Case for
Optimization curve is between cost, time and C-T curve. Economy Grid Architecture for Service-Oriented
The essence of C-T strategy is cost optimal strategy, it Grid Computing, Proceedings of the International
have little difference with cost optimal scheduling. Where Parallel and Distributed Processing Symposium: 10th
Deadline and Cost Optimization mixed the time and cost IEEE International Heterogeneous Computing
factors, by adjusting the weights, we easily adjust Workshop (HCW 2001), April 23, 2001, San
preference; hence it is dynamic and overall. Francisco, California, USA, IEEE CS Press, USA,
As the resource number increasing in the proportional 2001
share system, it can be find in Figure1 that the average [4] D. Ferguson, C. Nikolaou, J. Sairamesh, and Y.
finish time of job is decreasing, and the average cost of Yemini, Economic Models for Allocating Resources
execute job is also has a slight reduce. Moreover, the in Computer Systems, In Market-based Control: A
average cost becomes stable, i.e., when more resource Paradigm for Distributed Resource Allocation, World
adds into system, it can not reduce the average cost of the Scientific Press, Singapore, 1996
whole system any more. [5] K.M.Sim, From market-driven agents to market-
oriented Grids, ACM SIGECOM: E-Commerce
Exchange, vol.5, no.2, pp.45–53, Nov.2004.
5. Conclusion and future work [6] Kelly, F., A. Maulloo, and D. Tan: 1998, `Rate
control for communication networks: shadow prices,
This work explores issues of framework about proportional fairness and stability'. Journal of the
economic models and different schedule algorithms in Operations Research Society 49(3), 237-252.
Nimrod-G. Based on the proportional share system, we [7] S. J. Regehr. Some guidelines for proportional share
present a deadline and cost optimization schedule CPU scheduling in general-purpose operating
strategy, and extend schedule method of Nimrod-G in to systems. In Work in progress session of the 22nd
proportional share system. In contrast to cost optimal and IEEE Real-Time Systems Symposium (RTSS 2001),
time optimal, deadline and cost optimization strategy London, UK, Dec. 2001.
considers a coefficient weight factor. In contrast with the [8] I. Stoica, H. Abdel-Wahab, K. Jeffay, S. K. Baruah,
traditional schedule strategy, deadline and cost J. E. Gehrke, and C. G. Plaxton. A proportional share
optimization is easier to realized and more dynamic for resource allocation algorithm for real-time, time-
weight adjustment. shared systems. In IEEE Real-Time Systems
Consider the problem of software brokers being used symposium, December 1996.
as proxies for the procurement of computational Grid
resources and applications requirement. We investigate a
time optimization algorithm, a cost optimization
algorithm, a deadline and cost optimization algorithm
50