University College London, Dept. of Electronic & Electrical. Engineering, Torrington Place, London WC1E 7JE, UK ... types. The research towards RoD from the network point of view has yet been significantly ... background and related work, mainly concerning the QoS .... to active nodes using reserved socket numbers.
Network Engineering towards Efficient Resource on-Demand in Grid Computing Kun Yang, Xin Guo, Alex Galis
Bo Yang, Dayou Liu
University College London, Dept. of Electronic & Electrical Engineering, Torrington Place, London WC1E 7JE, UK
Jilin University, School of Computer Science, No. 10 QianWei Road, Changchun, JiLin, 130012. P. R. China
{kyang | xguo | agalis}@ee.ucl.ac.uk
{yangbo | dyliu}@mail.jlu.edu.cn
Abstract: The essence of Grid Computing is to provide efficient Resource on Demand (RoD). This paper addresses this challenge from the perspective of network, the living platform of Grid, by providing effective Quality of Service (QoS) mechanisms (both IntServ and DiffServ) inside the Grid networking environment. Particularly, the efficiency of this QoS mechanism is maximized by policy-based management taking care of the flexible control of QoS parameters/components and active networks technology looking after the fast delivery of various QoS configurations. An early experiment exemplified the current implementation status.
support for advance reservation and heterogeneous resource types. The research towards RoD from the network point of view has yet been significantly taken into consideration. But network, as the transporting media for Grid services, is critical to guarantee fully efficient RoD. Obviously, the bad quality of service (QoS) in the networks can significantly obstruct the efficient provisioning of RoD. This paper is trying to cope with these challenges from the network engineering point of view by providing efficient QoS mechanism inside the Grid networking environment. Particularly, in order to maximally gain efficiency for RoD, policy-based management method is exploited as a means to enhance or modify the functionalities of policy-influenced QoS components in a more flexible and automated way; while active network technology is utilized for the fast delivery of various configurations
Keywords: Grid Computing, Resource on Demand (RoD), Quality of Service (QoS), Active Networks (AN), Policy-based Management (PBM)
I.
INTRODUCTION
Grid computing is distinguished from conventional distributed computing or cluster computing by its focus on large-scale resource sharing, innovative applications, under the environment of widely-connected network, typically the Internet [1]. As it was in cluster computer, transparent access of resources across the network is still the main concern of Grid computing, but with more complexity due to the large scale of networks, the living platform of Grid computing. Deriving from the motivation of Grid computing, the essence of Grid Computing can be summarized as Resource on-Demand (RoD), i.e., transparently providing the Grid resources needed by Grid applications or services without or with little time delay. Compared to the traditional cluster computing or load balancing which typically happened in local area networks, efficiency turns out to be a big concern since resources in Grid computing environment are likely to be naturally distributed widely around the globe, hence more potential delay is expected. Efficiency means the fast delivery of RoD requirement. Most current researches with same goal of providing efficient use of Grid resources are carried out from the resources themselves’ point of view. One good example is Globus Architecture for Reservation and Allocation (GARA) system [2], which extends the Globus [3] resource management architecture by providing new features such as
The content of the paper is structured as follows. Based on the discussion in this section, Section 2 presents background and related work, mainly concerning the QoS issues, policy-based management and active networks. Then system architecture and its components for enabling the RoD are detailed in Section 3. Section 4 shows a first experiment demonstrating the feasibility of this network engineering method. A brief conclusion and future work are described in Section 5.
II. BACKGROUND & RELATED WORK A.
Resource on-demand: Essence of Grid Computing Grid computing addresses several trends that are challenging IT to improve the efficiency of their distributed systems today, such as increasing time-to-market pressure, ever-growing data sets, and user-friendly use of network resources. The concept behind grid computing is actually quite simple and sensible. A "Grid" consists of computational systems, software, and resources such as CPU cycles, massive storage, or scientific instruments. Grid computing supporting software such as Globus [3] pools these resources together into virtual single systems and makes it possible for anyone on the network to access them transparently and userfriendly, as depicted in Figure 1. Simply speaking, Grid computing provides resources on demand (RoD).
Grid Computing Virtual Grid Resources Grid Resource Management
Efficiency Resource on-Demand (RoD)
(physical Grid resources)
Policy-based Policy-based Management Management
Quality Qualityofof Service Service(QoS) (QoS) Technology Technology Active Active Networks Networks
Figure 1: Efficient Resource on Demand in Grid Computing Environment
In order to provide efficient RoD, mechanisms from only Grid Resource Management are not enough. Since all the Grid transactions regarding resource usage take place via large-scaled networks, quality guarantees from networks have to be considered. In order to make the RoD a reality, this paper focuses on the efficient QoS technology, as depicted in the right part of Figure 1. B.
Quality of Service: DiffServ and IntServ Our approach of addressing the QoS problem exploits both resource reservation and Differentiated Service (DiffServ) as the fundamental guideline. DiffServ is widely seen as the promising way to deal with data flows that have different time-bound constraints. DiffServ is expected to scale across the Internet. But DiffServ itself can’t solve the end-to-end QoS problem, thus Integrated Services (IntServ) is adopted to deal with QoS problems at ends. IntServ and DiffServ are the well-established paradigms proposed by IETF to achieve end-to-end IP QoS for a wide variety of network applications including Grid application [4]. However, IntServ is more appropriate for small private networks than the core of the Internet. On the other hand, DiffServ has shown well scalability to large networks. Actually, they compliment each other very well. C.
Policy-based Management for Grid QoS End-to-end QoS using both DiffServ and IntServ can be very complex in the Grid computing environment, and this raises the increasing requirement for the management of QoS itself. Policy-based management (PBM) is a good candidate for such complex management environment. In comparison with previous traditional network management approaches, such as TMN or TINA-C, PBM offers a more flexible, customisable management solution that allows controlled elements to be configured or scheduled on the fly, for a specific requirement tailored for a customer [5]. P. Flegkas et al. has explored the applicability of PBM to the management of QoS and proved its feasibility [6]. But this work was carried out within the scope of IP networks therefore without
considering the Grid computing environment and its corresponding management requirement. On the other hand, most researches focusing on Grid management have yet paid much attention to PBM, with these typically represented by Condor-G system and Nimrod-G Grid resource broker. K. Yang, et al. [7] presented a policy-based Grid management architecture supervising the overall Grid management. This paper is regarded as a continuation of this architecture with emphasis on the QoS management as part of this Grid management architecture. D.
Active Network Technology for Grid QoS Apart from efficient and automated management of QoS at higher level by using PBM, the even more critical part for an efficient end-to-end QoS (therefore RoD) is the delivery of these management requirements, i.e., how to configure the routers to make it support required QoS parameters in a fast, dynamic and automated way. Active Networks technology was initiated for this mission. Active network transforms the store-and-forward network into store-compute-and-forward [7]. The innovation here is that packets are no longer passive but rather active in the sense that they carry executable code together with their data payload. This code is dispatched and executed at designated (active) nodes performing operations to change the current state of the node. Active network is distinguished from any other networking environment by the fact that it can be programmed. In this paper, this programmability is fully guided by the QoS management policies. While Active Network research has precisely tackled that problem domain in fixed and wireless network environments, the particular conditions of Grid networks have not sufficiently been taken into account. Based on the work done in [7], this paper contributes to the state-of-the-art of active Grid by exploiting the applicability of AN to Grid QoS to achieve the efficient RoD in Grid environment.
III. SYSTEM ARCHITECTURE FOR ENABLING EFFICIENT RESOURCE ON DEMAND Figure 2 shows the system architecture and its basic components for enabling efficient RoD in Grid computing environment by utilising both DiffServ and IntServ. PBM components are used to thread the other components from DiffServ, IntServ, active network and Globus together to form a fully integrated system. The implementation is based on the Grid supporting tool Globus. A.
Policy-based Management Components In order to deploy PBM technology, a standardization process should be followed to ensure the interoperability between equipment from different vendors and, furthermore, PBM systems themselves from different developers. The framework and policy information model defined by IETF
Policy Framework Group gains wider popularity and is adopted as the baseline for the PBM system used in this paper. Policy-based QoS Management Tool
LDAP XML: XML: DiffServ DiffServ Scheduler Scheduler
Signalling DiffServ PDP
BB
Admission Control Monitoring
Policy Repository
Request Scheduling Resource Scheduling
LDAP IntServ PDP Resource Reservation Monitoring
Route Management Admission Control
Grid Supporting Environment (Globus) code DB
COPS
Active Packets IPv6 Router + Active Engine (PEP)
SNMP, CMIP
WLAN Centre/ Access Points
Active Packets
(PEP)
Figure 2: System Architecture for providing Efficient RoD
As illustrated in Figure 2 from top down, the PBM system for Grid QoS management mainly includes four components: policy management tool, policy repository, Policy Decision Point (PDP) and Policy Enforcement Point (PEP). Policy management tool serves as a policy creation environment for the administrator to define/edit/view QoS policies in a highlevel declarative language. After validation, new or updated policies are translated into a kind of object oriented representation or so-called information objects and stored in the policy repository. The policy repository is used for the storage of policies in the form of LDAP (Lightweight Directory Access Protocol) directory. Once the new or updated policy is stored, signaling information is sent to the corresponding PDP, which then retrieves the policy and enforces it on PEP. There is a need of a transport protocol for communication between PDP and PEP so that PDP can send policy rules or configuration information to PEP, or read configuration and state information from the device. A wide range of protocols can be used here, such as SNMP, CMIP or COPS (Common Open Policy Service), among which COPS is becoming the standard. An object oriented information model has been designed to represent the QoS policies, based on the IETF PCIM (Policy Core Information Model) [8]. Policies are further represented by XML due to XML’s built-in syntax check and its portability across the heterogeneous platforms. DiffServ PDP and IntServ PDP are integrated with Grid supporting tool, Globus, to support the potential inter-PDP
communication among QoS PDP and other Grid specific PDPs like storage scheduling PDP. Various services for Grid application can be introduced by defining new policies, e.g., to apply a new DiffServ shaper, or modifying existing policies. Then the Java classes for fulfilling these policies, which abide by class hierarchy and naming rules of policy information model developed within this system, can be instantiated by QoS PDPs according to these policies. B.
IntServ PDP and DiffServ PDP A typical Grid QoS framework includes a DiffServ region interconnecting two IntServ/RSVP regions at each side where Grid applications and Grid resources are connected respectively. RoD applications use IntServ/RSVP to communicate the quantitative QoS requirements. RSVP signaling messages travel end-to-end between sender and receiver to support reservations in the IntServ network regions. When crossing DiffServ region, these end-to-end RSVP messages are tunneled. IntServ/RSVP signaling requests specify an IntServ service type and a set of quantitative parameters known as a “flowspec”. Admission control component for DiffServ network regions maps IntServ service types to a corresponding DiffServ service level (DSCP or PHB) that can reasonably extend the IntServ service type requested by RoD applications. The admission control component, which is part of DiffServ PDP can then approve or reject resource requests based on the capacity available in the DiffServ network region at the mapped service level. More information concerning the mapping between IntServ and DiffServ parameters can be found from MANTRIP deliverables [9]. C.
ABLE Active Engine for fast QoS Delivery Active node, the core of the active network, is based on ABLE. ABLE [10] is an active network architecture that primarily addresses the network management challenges. Its major component is the active engine that is attached to any IP router to form an active node. ABLE can easily be deployed in Grid networks based on standards, i.e., the mobile code is written in Java, and it is encapsulated together with data using the standard ANEP (Active Network Encapsulation Protocol) headers over UDP. The Java bytecodes configuring the routers to support QoS are stored in the code database and can be downloaded by QoS PDP to the policy-based management station located in the ISP domain to which Grid application connects. Then this Java class is encapsulated into ABLE packets and is sent to active nodes using reserved socket numbers. After the packet arrives at an active node, the ABLE active engine (actually divertor) collects it, opens a file and starts writing the Java class contained in the packet to it. It then invokes the
SessionManager which forks a loader for a Java Virtual Machine (JVM). The JVM starts by invoking a SecurityManager which in turn loads the class being written to the file. Once the class is loaded, the active session begins and the loaded Java class starts running to reconfigure the router to make it support the instantiated QoS requirements as produced by both Grid application and policy-based management system. One big advantage with active networking is that functions can be dynamically provisioned. This implies that they can be installed, de-installed, configured or even programmed on-demand. As such, edge routers and core routers are enhanced with AN execution environment (EE) that fulfils the functionality of active networks according to the policies given by administrator or generated automatically in response to the RoD requirement. Active routers are used for many purposes as far as Grid QoS is concerned, such as, fast modification of admission or congestion control algorithms, dynamic deployment of customized or advanced reservation protocols and Per-HopBehaviors (PHB), etc.
IV. EARLY EXPERIMENT Based on the system architecture given above, this section presents an early experiment to evaluate this architecture, i.e., policy-based active QoS provisioning in the Grid environment, as shown in Figure 3. Both resource reservation (IntServ) and DiffServ are used in order to obtain the best of both and avoid the worst of either. Policy-based Active QoS Mgmt. Station XML
Grid App.
ER1
IntServ Region
BR1
XML
XML
BR2
DiffServ Region
V. CONCLUSIONS AND FUTURE WORK While most of the work on providing efficient resource sharing in Grid computing has focused on the managing and scheduling of physical resources themselves, our work addresses this challenge from the network engineering point of view. From the first experiment demonstrated above, we can see that after administrator inputted the requirement in response to the RoD requirement from Grid application, the whole procedure processed automatically. The whole system also scaled to the changing of the application requirement automatically. The integration of the Grid QoS with other resource management mechanisms such as GRAM (Grid Resource Allocation and Management) to get a complete end-to-end RoD of efficiency will be the main concentration of future work. The integration of this policy-based QoS management system with other Grid management system is also an interesting research topic; while the deployment of the above mechanism in a large scale is obviously a very big challenge.
REFERENCES [1]
AN EE Active Packet
Code DB
scheduler) was downloaded from code database on to edge router 1 (ER1, also served as domain management station) and was encapsulated into ABLE active packets. These active packets travelled along the route to configure routers so as to make these routers QoS-sensitive. Finally, the quality of service was guaranteed in the network and RoD was ensured given the necessary configuration on both Grid application and Grid resource had been carried out.
ER2
IntServ Region
Grid Resource
Figure 3: Active Grid QoS Scenario
Grid administrator, or software on his behalf, used Active QoS Management Station to manage the underlying network environment (including one DiffServ Region that is connected to IntServ Region at each end) by giving policies. After being notified the new change of policies, QoS PDPs residing also on QoS management station in this scenario retrieved from policy repository the new policies, based on which QoS PDPs made decision and a set of policies for fulfilling the task required by Grid administrator or Grid application were selected or generated. These policies were further translated into XML files and transported to active network EE which also served as PEP to enforce the policies in term of activating the certain active packets processing. Then, guided by QoS PDP, code (e.g., a new DiffServ
I. Foster, et al. "The Anatomy of the Grid: Enabling Scalable Virtual Organizations". Intl. J. Supercomputer Applications, 15(3), Sage Publications, USA, 2001. [2] I. Foster, et al. “A distributed resource management architecture that supports advance reservations and co-Allocation,” Proc. Of International Workshop on Quality of Service, pp. 27-36, 1999. [3] I. Foster and C. Kesselman “Globus: A Toolkit-Based Grid Architecture”. The Grid: Blueprint for a New Computing Infrastructure. pp, 259-278. Morgan Kaufmann, 1999. [4] Y. Bernet, et. al. “Interoperation of RSVP/Intserv and Diffserv networks”. IETF Internet Draft, June 1999. < draft-ietf-issll-Diffserv-rsvp-02.txt> [5] M. Sloman. “Policy Driven Management For Distributed Systems,” J. of Network and System Management, 2(4): 333-360, Dec. 1994. [6] P. Flegkas, P. Trimintzios, and G. Pavlou. “A Policy-Based Quality of Service Management System for IP DiffServ Networks”. IEEE Network Magazine, pp. 50-56. March/April 2002. [7] K. Yang, A. Galis, C. Todd. “A Policy-based Active Grid Management Architecture”. Proceedings of 10th IEEE International Conference on Networks (ICON02), pp. 243-248, IEEE Press. August 2002. [8] J. Strassner, E. Ellesson, and B. Moore. “Policy Framework Core Information Model”. IETF Policy WG, Internet Draft, May, 1999. [9] European Union Information Society Technology (IST) Project MANTRIP web site: http://www.solinet-research.com/mantrip/ [10] D. Raz and Y. Shavitt. “Active Networks for Efficient Distributed Network Management”. IEEE Communications Magazine, 38(3):138143, March 2000.
