Configuration Management Policy in QoS-constrained Grid Networks Hyewon Song1, Chan-Hyun Youn1, Chang-Hee Han1, Youngjoo Han1, San-Jin Jeong2, Jae-Hoon Nah3 1
School of Engineering, Information and Communications University (ICU) 103-6 Munji-dong, Yooseong-gu, Daejeon, 305-714, Korea {hwsong, chyoun, chyoun}@icu.ac.kr 2 Mobile Telecommunication Research Division, Electronics and Telecommunications Research Institute (ETRI) 161 Gajeong-dong, Yuseong-gu, Daejeon, 305-700, KOREA
[email protected] 3 Information Security Research Division, Electronics and Telecommunications Research Institute (ETRI) 161 Gajeong-dong, Yuseong-gu, Daejeon, 305-700, KOREA
[email protected]
Abstract. In Grid service, resource management is important to support capability for good quality and efficiency for the computing and storage service. In order to provide this resource management which has low complexity, high performance and high throughput with guaranteeing QoS, the well defined network resource management architecture and scheme have to be considered. Thus, we propose the integrated grid resource management architecture based on QoS-constraint policy and we derive cost-sensitivity policy enforcement procedure which can be applied to this architecture in Grid network. Finally, the results show that the proposed scheme outperforms the conventional scheme in the network performance such as a blocking rate of resource request messages and the operation cost.
1 Introduction The global growth of Grid computing means the growth in application part reflecting the needs of users. The traditional Grid computing (Computational, Data, Access) [1]-[3] is more specialized, and consist of very various and complex grid services. All the more Resource management need to correspond with the need of users or application and policy based resource management system is one of the most promising candidates for solving requirement of resource management. For the reliable management of physically distributed and heterogeneous resources, policy-based management [4] has been suggested. The connection between lower network and the resource management system is mandatory for such a grid policybased resource management system. In other words, it is requisite to make the re-
source management system abstract because it is hard to manage heterogeneous lower management system managed by local policy of different subject like a grid plane. However, this [4] isn’t consider about network resource management of lower network layer such as L3/L2/L1. Actually, since the high performance and throughput and QoS-constraint service requested with a specific Service Level Agreement (SLA) can be supported throughout Grid service networks, the network resource management is also considered importantly. For this network resource management, there are many studies [8]-[10]. Especially, for high throughput issue, optical network is mainly considered. [10] Moreover, as mentioned in [11], optical network with using GMPLS give more efficient management and control scheme for Grid and network resource management in Grid service networks. In this paper, we propose the integrated grid resource management system architecture based on Quality of Service (QoS) constraint configuration management policy in Grid networks. This QoS-constraint configuration management policy is based on the cost sensitivity using the cost according to providing network resources. Thus, we define the cost function and cost sensitivity function for guaranteeing the QoS in using network resources in this paper, and then, we derive the configuration management policy rule. Also, we propose the policy enforcement process using this policy rule in proposed grid resource management system architecture which is possible to implement in Grid networks. Finally, through the theoretical and experimental analysis, we can show that the proposed scheme outperforms the conventional scheme in network performance and cost performance.
2 Related Works The Policy Quorum based Resource Management Architecture is a kind of grid resource broker and scheduler that manage resources and jobs in the virtual Grid environment based on Globus Core.[4] Policy quorum, be generated finally, represents the collection of the resources that is adapted by individual policies according to private request. Thus a user is satisfied with QoS by Policy Quorum. And several papers have showed the resource reconfiguration algorithm based on temporal execution time estimation method. Resource reconfiguration performs the reshuffling of the current available resource set for maintaining the quality level of the resources. However there is no consideration of the cost when reallocation job. The policy-based Grid management middleware sits on the top of Gird Supporting Environment and serves as the first access point for Gird administrator, or software on its behalf, to configure and manage intra-domain and inter-domain Grid resources. In order to deploy Policy-based management technology in Grid architecture, a policy-based management standard architecture of the Internet Engineering Task Force (IETF) can be used. These policy based management (PBM) system is originated from network management groups to reduce the complexity of management in terms of QoS guarantees. It is suitable to complex management environment such as large scale heterogeneous management in Grid networks. Quorum is a collection of the elements which always intersects. It has been used to maintain the consistency of replicated data, mutual exclusion. In PQRM the intersection property of Quorum is
used to make complex resource elements abstracted Grid resource which satisfies various QoS requirements. The main result of PQRM is that QoS-constraint policy has an advanced method to satisfy QoS requirements requested by applications or services [4]. Although this PQRM has many benefits for Grid resource management, Grid management considered about network resource and service isn’t considered like many other Grid resource management architecture. Actually, it is important to consider these network issues in general Grid resource management for supporting better QoS of Grid service. These issues are proposed in GGF, and then, now many researches are going on. [8]-[10] Most of them have an overlay model in a point of architecture view. However, since the integrated architecture is more efficient ultimately, we focus on an integrated model. In this architecture, the GMPLS is supposed for management of network resource in L1/L2/L3. Control and management capability of GMPLS is important to create new services by enhancing the controllability of optical networks. As mentioned in [11], geographically-distributed computing applications for science and enterprise are emerging services for providers, and the optical network is expected to be involved in such the Grid computing applications by using GMPLS protocol. This GMPLS can accept dynamical application specific requirement such as network delay and fault recover time which is requested by geographically-distributed application. Thus, [11] consider a network resource management function block based on management and control scheme in GMPLS networks. Throughout this method, this architecture can get the integrated framework between Grid resource management and network resource management. Similar to this architecture, in this paper, we propose more detailed and modified architecture and procedure in order to manage network resource cost-effectively and consider QoS guaranteed by providing efficient network enforcement scheme based on this architecture. That is, in this paper, we propose the policy based Grid resource management architecture and the cost sensitivity based policy enforcement scheme for the QoS constraint Grid management based on PQRM in proposed architecture.
3 Policy Enforcement Scheme based on Cost Sensitivity When a user requests some jobs with a specific SLA, the Grid resource manager select proper resources and reserve selected resources. After that, the job is distributed by a job scheduler and then, is executed in distributed resources. In order to select and reserve proper resources, the policy based resource management scheme using QoS constraint Quorum can be considered in [4]. As mentioned in a previous part, the reservation of network resource can be considered as an important performance factor when the QoS guaranteed grid service is considered. Thus, in this paper, we consider Grid resource management architecture with network enforcement policy based on cost sensitivity referring the PQRM architecture in [4]. We propose the policy enforcement scheme based on cost sensitivity (PES-CS) in this paper. In order to provide our proposed PES-CS, we define the cost sensitivity based network policy enforcement engine (CS-NPEE), which interacts with Grid resource management engine based on QoS constraint Quorum. Namely, throughout this CS-NPEE, the
policy determined by Grid resource management engine can enforce to network resources directly. In this process, the proposed PES-CS can be applied in order to select and configure network resources for providing the Grid service when a specific job is requested. In this section, we define operation cost function and cost sensitivity function. Also, from this function, we derive the policy scheme for the PES-CS and propose the basic procedure for this PES-CS. 3.1 Cost Sensitivity For network enforcement, we can consider the cost sensitivity based policy enforcement scheme. This scheme is based on network status and cost function when a network node is under a contention because of resource competition. In this section, we define some network cost function, and then, derive cost sensitivity based policy enforcement scheme. For underlying network, we suppose optical networks based on GMPLS for its control and management process. For this network, we can consider network status such as a status under guaranteeing QoS, NSQoS and a status under contention but guaranteeing tolerable QoS by providing alternative path, NS alt . [5] Using this network status, we can define a function x (t ) , which means the binary value according to those two statuses, NS QoS and NS alt , as follows. x (t ) =
⎧1, ⎨ ⎩0,
when NS = NS QoS when NS = NS alt
(1)
Also, we can derive an operation cost model under various statuses in [5]. Using Eq.(1) and this cost function, the total cost function when providing the service throughout the path between the source s and the destination d is derived as follows, sd sd Fsd (t ) = x(t ) ⋅ CQoS (t ) + (1 − x(t )) ⋅ (CQoS (t ) + Calt (t ))
(2)
sd
where CQoS (t ) is a cost based on a Deterministic Effective Bandwidth (DEB) concept in order to guaranteeing QoS by the SLA (Service Level Agreement) of a Grid service requested by clients and Calt (t ) is a additional cost of providing a alternate path in order to guarantee the QoS under contention situation in underlying networks [5][6]. This total cost function means the cost in order to provide the path which guarantees the QoS. When the data for supporting a Grid service is transmitted from source s to destination d through a network controlled by GMPLS, if the bandwidth for guaranteeing the QoS constraint of this data is assigned, only the cost based on DEB is considered for the total cost function. However, if the bandwidth for guaranteeing the QoS constraint of this data can’t be assigned by the reason such as contention of resource or blocking status, the alternate path is needed to guarantee the QoS. Thus, in this case, the total
cost function is represented by the sum of the cost based on DEB and the cost resulted from providing the alternate path. Moreover, when it is no meaning that guarantees the QoS because of a continuous increment of the operation cost, the total cost is represented by the penalty cost applied differently according to the service type because of dropping the data [5]. However, as mentioned in previous section, we assume that the value in the drop status NSbe in which the required QoS by SLA is not guaranteed is not considered since our proposed configuration management scheme relates to guarantee the QoS. When the total cost from Eq. (2) is considered between source and destination, to increase this cost means that the cost for guaranteeing the required QoS increase, especially when the network status changes such as the case of providing the alternate path. When the amount of traffic per each channel is expected by the data scanning in previous section, we can define the sensitivity of the total cost from Eq. (2) as follows: ζ Fsd =
∂Fsd (t )
(3)
sd (t ) ∂CQoS sd
when the Fsd (t ) is given by Eq. (2) and CQoS (t ) means a cost according to DEB cost [5], respectively. ζ F means the variance of total cost according to the variance of the cost based on DEB when traffic flows throughout the path between source s and destination d. When the sensitivity of the total cost is given by Eq. (3), we can derive the sensitivity according to the network status using the total cost function, F. sd
ζ Fsd = x(t ) + (1 − x(t )) ⋅ (1 +
∂Calt (t ) ) sd ∂CQoS (t )
(4)
sd
When we consider the sensitivity according to CQoS (t ) , the sensitivity value of the sd
total cost F is dominant to the variation value of both Calt (t ) . Therefore, when the node is under a contention situation, this is, x (t ) = 0 , ζ F dominantly depends on sd
the term, ∆ = ∂C (t ) / ∂C (t ) , which represents the variation of the cost for providsd
alt
QoS
ing the alternate path according to the cost based on DEB. When the alternate path is used at contention situation in Grid networks, the cost for providing this alternate path occurs. This cost increases when the number of hops in provided alternate path increases [5]. In high channel utilization of overall network, selected alternate path includes many hops since the high channel utilization means that most of channels has the traffic load which is closer to the boundary so that most of nodes is under the contention situation. Therefore, the value of ∆ can have a positive value because the cost for the alternate path increases. However, if the utilization of channels in overall network is closer the boundary, it becomes hard to reserve the resource for the data. Accordingly, the selected alternate path has to include more
hops. This increment of the number of hops causes the increment of the cost. Thus, the value of ∆ increases, so that the point in which this value exceeds the upper bound occurs. This upper bound is given by SLA. By this upper boundary, to provide the alternate path is determined. When the sensitivity of total cost, ζ F , has the sd
boundary by ζ F ≤ 1 + ∆ , the value exceeding this boundary has no meaning in the network operation cost point of view, so that it needs not to provide the alternate path in this case. sd
3.2 Policy Enforcement Scheme based on Cost Sensitivity with PQRM In this section, we describe the PES-CS procedure and PES-CS algorithm which can apply to Grid networks using proposed Policy based Resource Management Architecture using PQRM.
Fig. 1. The basic procedure for policy enforcement scheme based on cost sensitivity with the Policy based Resource Management using PQRM
Fig. 1 shows the basic procedure of this PES-CS for applying to Grid over GMPLS networks. A Network Status Information Base (NSIB) is a kind of data base for collected network status information and a Network Status Table (NST) is an updatable table for selecting network resources. As shown in Fig. 1, the network status information is saved and updated by monitoring procedure in NSIB, and NST in CS-NPEE is created and updated by this information of the NSIB periodically. Moreover, whenever this NST is updated, this updated NST is enforced in that network throughout the CS-NPEE. Moreover, as shown in a top part of Fig. 1, this update and enforcement process interact with QoS Quorum calculated and selected by Policy based Resource Management using PQRM. If the QoS Quorum, which matches with a specific SLA requested by user, is selected by QoS Quorum Generator, this information is reflected in a network enforcement process. Also, during generation of QoS Quorum in QoS Quorum Generator, NST and network status information in NSIB are used in this process like an upper side of Fig. 1.
In the CS-NPEE procedure, the value of NST reflects the network status information which is represented by three network statuses. [5] That is, the value of NST, NS , changes according to the network status, and is then updated by the proposed ij
decision algorithm. When it is assumed that a job with a specific SLA is requested and the contention information in nodes is given for the update of NST, we propose the decision algorithm in order to provide proper network resources by PES-CS. For these algorithms or procedures, we define some factors in this section. We have the condition factor by the number of hops, Qh , as follows:
⎧1
if H sc − H cd ≥ 0
⎩0
otherwise
sd
Qh = ⎨
where H
sd sc
sd
.
(5)
sd
and H are the number of passed nodes before a current nodes in this cd
path and the number of remaining nodes after a current node in this path, respectively [5]. If the current node is closer to destination d, the value of Qh is one, otherwise, the value of Qh is zero. Also we can obtain the other condition factor, Qδ , from the sd
F
boundary in a previous section and it is defined as follows: Q
δF
sd
=
⎧1 ⎨ ⎩0
if ζ F ≤ 1 + ∆ and x (t ) = 0 sd
.
(6)
otherwise
where ∆ = ∂C (t ) / ∂C (t ) and x (t ) is value according to network statuses. When sd
alt
DEB
the decision factor is represented by a sum of above two condition factors, the combined threshold check function can then be stated as C ALT =
⎧1 ⎨ ⎩0
if Qt
=w
δ
+ wh
.
(7)
otherwise
Using previous parameters, we can determine the threshold check function. When the current node between source and destination is under a contention situation, if the node that is closer to destination d and the value of ζ F , which represents the sensisd
tivity of the total operation cost, is within the tolerable range, the combined threshold check function C is one, so that the node makes a decision to deflect the alternate ALT
path. Otherwise, C is zero, so that the node makes a decision to drop the data packet. When information is obtained from NST and NSIB, the threshold check function is determined. Finally, when the current node is under the contention situation, the node makes a decision whether the data is to be deflected to an alternate path or dropped according to the threshold check function, C . As mentioned in a previous part, the procedure ALT
ALT
for updating and enforcement in CS-NPEE relates to the upper part – the Policy based Resource Management using PQRM. Fig. 2 shows the flow chart used in CS-NPEE. Job (or Service) Request
⎧1, Qth = ⎨ ⎩0,
C fr (t ) ≤ r * otherwise
Qth = 1 ?
where C fr (t ) means the additional cost with historical failure rate.
No
Single Resource Allocation Table according to SLA type of requested job is selected.
Yes Multiple Resource Allocation Table according to SLA type of requested job is selected.
Compare Algorithm
Satisfy?
No
Waiting
Yes
No
Contention ?
Enforce determined QoS ARQ to Network PEP.
Yes
NS ij = 1 ?
No
Drop the packet (Request refuse)
Yes
Deflection into alternate LSP
Yes
Available alternate LSP ?
Determine one of following actions: 1. Create New ARQ from RQ. 2. Renegotiation with user
Using QoS Quorum information, we can provide this algorithm like this algorithm/
No
Drop the packet (Request refuse)
Fig. 2. The Policy Enforcement Scheme based on Cost Sensitivity (PES-CS) procedure with QoS Quorum information by the Policy based Resource Management using PQRM
This procedure is applied by the CS-NPEE when the job with a specific SLA is requested and the resource scheduling is performed by the Grid resource management engine in the Policy based Resource Management using PQRM.
4 Simulation and Results In order to evaluate the performance of the proposed policy enforcement scheme, a simulation model is developed. We use the Grid network underlying optical network based on GMPLS in our simulation. The data sources were individually simulated with the on-off model. The simulation is carried out using a 14-node NSFNET network topology. The transmission rate is 10 Gb/s, the switching time is 10 us, and the data header processing time at each node is 2.5 us. The primary paths are computed using the shortest-path routing algorithm, while the alternate paths are the linkdisjoint next shortest paths for all node pairs. [5] Fig 3 shows the results from this simulation. The basic mechanism of proposed PES-CS is different from schemes in general optical network. As mentioned in previous parts, when the node is under the contention situation, the alternate path is provided by threshold value. On the other hand, the conventional general schemes provide retransmitting service at that time. This affects blocking rate in Grid networks. As shown in Fig 3, the blocking rate of the proposed scheme decreases about average 23.679% compared with the general scheme. Moreover, in order to evaluate the PES-CS scheme in terms of cost, we consider the traffic source that is leaky bucket constrained with an additional constraint in the peak rate based on [7]. We consider 50 different network status tables according to
randomly generated traffic patterns under the given conditions. We assume an interval among incoming traffic scenarios is a monitoring interval. For each scenario, we compare the values of total operation cost function between the existing general scheme and the proposed PES-CS. For a comparison, the upper boundary for PES-CS, 1 + ∆ , is assumed to be 100. In the case of existing general scheme, the total cost is about 4 times that of the proposed policy decision in terms of average total cost. This means that existing general scheme provides an alternate path in spite of high cost value. In addition, from the point of view of variation, the cost of existing general scheme fluctuates widely as shown in (a) of Fig 4. Also, (b) of Fig 4 shows that most of the sensitivity values in the case of PES-CS are constant, at 100. 1.E+00
Optical Burst Blocking Rate
1.E-01 1.E-02 1.E-03 1.E-04 1.E-05 1.E-06 1.E-07 1.E-08 0.6
0.65
0.7
0.75
0.8
0.85
0.9
Traffic Load general scheme
proposed PES-CS
Fig. 3. Blocking rate comparison of general scheme and proposed PES-CS
160
700
140
600 The sensitivity of total cost
total operation cost
120 100 80 60 40 20
500 400 300 200 100
0 1
3
5
7
9
11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 change of traffic pattern
0 1
3
5
7
9
11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 change of traffic pattern
under proposed PES-SC
under proposed PES-CS
under conventional scheme
under conventional scheme
(a)
(b)
Fig. 4. (a) The total operation cost comparison, (b) The sensitivity comparison
5 Conclusion In this paper, we proposed policy based Grid resource management architecture based on existing PQRM and procedure in order to manage network resource and consider QoS guaranteed by providing efficient network enforcement scheme based on this architecture. For this policy enforcement scheme for the modified policy
based Grid resource management architecture, we developed a modified operation cost model according to the network status information changed by guaranteed QoS. In addition, using the bounded range of the sensitivity of this cost, we proposed a network status decision algorithm, and developed policy decision criteria for policy enforcement in Grid networks by providing an alternate path. As shown in the comparison of the cost performance between our proposed scheme and conventional schemes, our scheme is performed under a stable state. As well, in comparing the blocking rate between our proposed scheme and conventional schemes, ours has good performance in terms of blocking rate. Finally, by using the bounded range of the sensitivity of the total operation cost, our proposed scheme has a reducing effect of about 24% in terms of total operation cost.
Acknowledgement This research was supported in part by ITRC (Information Technology Research Center) and MIC (Ministry of Information and Communication) of Korean government.
References 1. Foster, I. and Kesselman, C. (eds.). “The Grid: Blueprint for a New Computing Infrastructure”. Morgan Kaufmann, 1999 2. Foster, I. and Kesselman, C. “The Anatomy of the Grid:Enabling Scalable Virtual Organizations”. Intl J. Supercomputer Applications, 2001 3. Czajkowski, K. et al. ”. Grid Information Services for Distributed Resource Sharing, 2001 4. Byung Sang Kim et al. “Policy Quorum based Resource Management Architecture in Grids”, IJCSNS 2005, Vol 5, No 8 5. H.W. Song, S.I. Lee, C.H. Youn, “Configuration Management Policy based on DEB Cost Model in OBS Networks”, ICAS-ICNS 2005, Oct. 2005. 6. C.H. Youn, H.W. Song, J.E. Keum, L. Zhang, B.H. Lee and E.B. Shim, “A Shared Buffer Constrained Topology Reconfiguration Scheme in Wavelength Routed Networks” INFORMATION 2004, Nov. 2004. 7. D. Banerjee and B. Mukherjee, “Wavelength routed Optical Networks: Linear Formulation, Resource Budget Tradeoffs and a Reconfiguration Study,” IEEE/ACM Transactions on Networking, Oct. 2000. 8. Masum Z. Hasan, et al, “Network Service Interfaces to Grid”, draft-ggf-masum-gridnetwork-0, GGF Draft, May 2004. 9. Doan B. Hoang, et al, “Grid Network Services”, draft-ggf-ghph-netwerv-2, GGF Draft, May 2005. 10. D. Simeonidou, R. Nejabati, et al, “Optical Network Infrastructure for Grid,”, draft-ggfghpn-opticalnets-1, GGF Draft, Mar. 2004 11. M. Hayashi, T. Miyamoto, T. Otani, H. Tanaka, A. Takefusa, et al, ”Managing and controlling GMPLS network resources for Grid applications,” OFC 2006