Special Issue Advanced Technologies Driving Dynamic Collaboration Featuring System Technologies

Polimatica: An Implementation of
Policy Automated Provisioning Grid
Foundation of Dynamic Collaboration
By Takashi KOJO,* Yoshiharu MAENO and Yoshiki SEO

Dynamic resource allocation is one of the critical characteristics required for the back-end systems
ABSTRACT
implementing Dynamic Collaboration. In this paper, we first discuss the requirements on such
systems and point out that the policy automation is the key technology. Then, scalable models for dynamic
resource brokering of complex large distributed systems are discussed. We implemented basic functions of the
models and successfully demonstrated feasibility of the model.

KEYWORDS Grid, Provisioning, Policy automation, Dynamic resource allocation, Dynamic Collaboration

1. INTRODUCTION concerned with those kinds of changes of system con-

figuration and deploys all the software required for
Dynamic Collaboration is supported by a collective the application. Although the technology reduces
of various types of applications. Some of them are many aspects of the administration overhead and re-
user access oriented interactive applications such as duces the total cost of the operations, it still requires
mobile applications or Web portals. Some of them significant portions of manual operations that pre-
require very high network traffics such as stream vent a serious cut down of the lead time for system
video applications. Some may be database oriented configuration changes.
while others may require computation-intensive Provisioning Grid discussed in this paper is an
tasks. Implementation of the dynamic collaboration emerging grid technology which focuses on policy-
systems requires various unique characteristics on based automation of dynamic resource allocation, re-
its back-end systems as well as the front-end and source sharing and utilization, as opposed to large
network. portion of the grid technology focused on IT resource
In such applications, for example, the system load sharing in the collaborative working environment
varies drastically from time to time as user access among multiple organizations. Provisioning grid en-
surges unpredictably, while some applications might ables fully automated change of its hardware and
have very periodic daily, weekly or monthly patterns software configuration as the system environment
of systems load. Respective applications require their varies. It can provide an optimal environment for
own hardware and software configurations based on each application for constantly changing the system
performance, security or other requirements. For its execution environment so that the application can
constantly changing system load, it would not be real- meet its own service level requirements from the us-
istic to assume maximum load on the system configu- ers, while the entire system can realize maximum
ration of each application. Although dynamic change resource utilization or very high availability.
of the system configuration and resource allocation as This paper describes the required characteristics
its change of system load is strongly required for of the provisioning grid, the internal models the sys-
entire system resource utilization, it usually includes tem is based on, and its implementation undergoing
many manual operations, time consuming and diffi- by NEC.
cult tasks. Conventional system administration
speed does not meet the constantly changing environ- 2. REQUIREMENTS
ment.
Provisioning is a technology based on business In this section, we discuss a variety of require-
process management that can pipeline the operations ments for the provisioning grid.

1) Manageable Resources
*System Platform Software Development Division To ensure a full range of optimization regarding
Internet Systems Research Laboratories the system configuration and resource allocation,

56 NEC J. of Adv. Tech., Jan. 2004

01-a10.p65 Page 56 2004/03/15, 13:37 Adobe PageMaker 6.5J/Win

 Polimatica: An Implementation of Policy Automated Provisioning Grid Foundation of Dynamic Collaboration

manageable resources of the grid have to include vari- have to be predefined by business processes and poli-
ous layers of the system resources from hardware to cies so that the application can adequately perform
application software. At the bottom layer, it has to its required task regardless of system environment
manage allocation of various types of server hard- changes.
ware, storage, firewall, load balancer and network The business process is a sequence of the activities
switches. Appropriate type and version of the operat- that can perform certain aspects of management. The
ing system has to be installed onto the allocated serv- processes are either infinite cycle of activities or a
ers. Middleware such as Web server, application finite sequence that is supposed to be started by
server or DBMS also need to be deployed as the appli- events.
cation required. For executing the application, the The policy is a pair of event and action. An event is
application program along with its initial data, con- any kind of phenomena in the system such as user
tents or database schema may also be needed on the access, change of the system load or any kind of ex-
allocated system configuration. ception or a familiars action. An action can be defined
All those entities are regarded as manageable re- as a business process which appropriately handles
sources. the event.

2) Resource Virtualization 4) Scalability

The resources should have their interfaces for Since the dynamic collaboration tends to extend
management. In the provisioning grid context, the over a wide range of groups, organizations or enter-
interface is virtualized into Web services which ex- prises, the back-end system which supports the col-
pose service ports for the management so that the laboration should be easily scaled out so that it can
provisioning grid middleware can manage the re- meet the service level requirements of the applica-
sources in a consistent manner as it manipulates tions. Aspects of the scalability include physical geo-
other grid services (Fig. 1). graphic distribution, logical number of organizations,
The scalability introduces heterogeneity in its na- performance to deal with size of users, organizations
ture. A large-scale system distributed among organi- and a variety of applications.
zations cannot assume homogeneity of the hardware,
operating systems or middleware. Standardization of As the system scales out to a large number of
the services is being discussed in several organiza- applications, the complexity of the systems behavior
tions. The most recent attempt is Web Services Dis- and nature also drastically increases. It is usually
tributed Management in OASIS Open. unrealistic for system designers to predict all the
possible system behavior and define the processes
3) Autonomy and Policy and policies that can manage an entire system appro-
An application should have its autonomy and a priately on any possible occasion. The grid has to
certain level of independence in terms of the resource provide adequate means to break down the problem
allocation and its behavior. Those characteristics into reasonable pieces. The processes and policies

Fig. 1 Provisioning layers.

NEC Journal of Advanced Technology, Vol. 1, No. 1 57

01-a10.p65 Page 57 2004/03/15, 13:37 Adobe PageMaker 6.5J/Win

 Special Issue Advanced Technologies Driving Dynamic Collaboration Featuring System Technologies

have to be broken down into local processes and poli- or public Internet space through a given bandwidth.
cies so that the system designer can deal with only The partition can be implemented in several different
the local behavior and nature of the system and can ways depending on the requirement of the network
break it down to a reasonable complexity of problems. security. The provisioning grid VO is implemented on
the partitioned layer two which realizes the virtually
3. MODELS most strict isolation from outside of the partition.
In the simple model, the resources inside the parti-
This section discusses various models on which the tion have equal access each other. However, to ad-
provisioning grid is based. dress realistic security requirements, we need to
implement a more sophisticated model in which clus-
1) Virtual Organization ter of resources have specific connectivity. Three-tier
VO (Virtual organization) is an isolated virtual Web configuration is an example of such connectivity.
work space for applications with which real organiza- We have introduced a template model for dynami-
tions can share (Fig. 2). The model is discussed in cally creating and managing the configuration.
OGSA (Open Grid Services Architecture). VO is both Three-tier Web configuration is a template on which
a security domain and resource allocation domain. As the system can allocate a number of appropriate
a security domain, the access from outside to the types of server resources as required.
resources inside the organization is totally prohibited
or controlled by policy of the VO. 3) Resource Brokering Model
As a resource allocation domain, VO owns re- Resource Broker is a mechanism for dynamically
sources allocated from resource pools in Real Organi- allocating resources owned by the ROs to VOs as they
zation. The resources are allocated to applications in are required. If the system provides a single central
the VO through appropriate mechanism and policy as brokering mechanism which handles all resource
it is needed. brokering from every RO to every VO, the system
scale out brokering policy would become extremely
2) Network Model complex and difficult if not impossible to design
Virtualization of the network devices is a critical appropriately (Fig. 3).
issue to minimize the complexity of the resource pro- To break down the design scope locally to each RO
visioning in these kinds of systems. VO can also be and VO and simplify the problem, we introduce a
extended to network virtualization. The provisioning push and pull models of the resource brokering (Fig.
grid implements VO model on top of collection of 4). The push brokering model is a resource provider
conventional network devices in which [Globus Tool centric model. Resources provided by a provider are
Kit 1][Globus Tool Kit 2] hides all the details of net- brokered among resource consumers. The provider is
work device set ups.[Globus Tool Kit 3] an RO and the consumer is a VO, in the grid system.
A VO owns isolated network partition. The parti- The broker has its brokering policy and the resource
tion is connected to the outside which is another VOs provider has its resource providing policy while the

Fig. 2 Real and virtual organizations.

58 NEC J. of Adv. Tech., Jan. 2004

01-a10.p65 Page 58 2004/03/15, 13:37 Adobe PageMaker 6.5J/Win

 Polimatica: An Implementation of Policy Automated Provisioning Grid Foundation of Dynamic Collaboration

consumers have their requesting policy. and action policy model is a simple but powerful ge-
The pull model is the counter model of the push neric policy model. It defines a set of events and
model. Resources required by a consumer are corresponding action to be taken.
brokered among available resources from multiple
providers. As with the push model, each component 4. IMPLEMENTATION
has its own policy.
A brokering model for multiple requesters and pro- 4.1 Functional Configuration
viders can be constructed by a combination of the A prototype system has been developed based on
push and pull model brokers (Fig. 5). The advantage the model discussed in the previous section. Figure 6
of the model is that its policies are localized to VOs or shows the functional configuration of the system. The
ROs so that the system designer can concentrate on a system consists of three layers. The VO management
particular VO or RO for the policy design. Another layer implements and manages VOs. The resource
advantage is that it is purely peer to peer architecture provider layer manages resources owned by ROs. The
and does not require central mechanism, so that it agent layer is for the grid middleware interfacing
can avoid scalability bottle neck for large systems with physical devices.
that create a large number of VOs.

4) Policy Model
The policy model is another key model for grid
implementation. A variety of the models both generic
and for specific application are discussed. The event

Fig. 3 Single broker model. Fig. 5 Combination of push and pull models.

Fig. 4 Push and pull models.

NEC Journal of Advanced Technology, Vol. 1, No. 1 59

01-a10.p65 Page 59 2004/03/15, 13:37 Adobe PageMaker 6.5J/Win

 Special Issue Advanced Technologies Driving Dynamic Collaboration Featuring System Technologies

The VO management layer consists of three grid gine controls the priority of policy so that, if events
services. Policy automation GS is for the top level are caused by error of a previous action, the recovery
process and policy implementing autonomic resource policy can override and start the action.
management of the VO. Monitor GS is for monitoring
the load of system components. Whenever the load 4.3 Resource Monitoring
level of a component is changed, it notifies the event Resource monitor GS monitor the resource status.
to the Policy Automation GS. Broker GS is a re- They periodically collect the status information from
quester side broker of the system resources. The cur- resource control GSs which correspond to respective
rent system manages a single RO. The system imple- managed entities. The entities are physical devices or
ments only pull model brokering. software. VO resource monitor and RO resource
The resource provider layer consists of Resource monitor are implemented on MDS (Monitoring and
Monitor GS, Resource Control GS and Mapping GS. Discovery Service). MDS is a service provided by glo-
Mapping GS virtualizes network device operations bus tool kit which provides directory services for grid
into network partitions owned by VOs. As the broker resources. MDS in the VO monitor caches resource
request to add or remove resources in the network information owned by the VO, while that in the RO
partition (VO), mapping GS interprets the request monitor stores resource information owned by the
and translates it into a series of commands to the RO. For implementing the VO based security, a VO
network devices. monitor is only allowed to access the resources owned
by the VO or free resources in the pool.
4.2 Policy Engine
The policy engine is the core of the policy automa- 4.4 Resource Pool
tion GS. The engine implements event and action The resource pool is also managed by the resource
type policies in which policy is a pair of event condi- monitor GS. The GS shows all the available resources
tions and an action definition. A policy is started if in RO to the inquiry by VOs. With current implemen-
one or more of the event conditions is satisfied and tation, the resources inquiries are serialized and the
requires corresponding action. resources are allocated to VOs on a first-come first-
The events have their assigned priority. The en- serve basis.

Fig. 6 Functional configuration of the system.

60 NEC J. of Adv. Tech., Jan. 2004

01-a10.p65 Page 60 2004/03/15, 13:37 Adobe PageMaker 6.5J/Win

 Polimatica: An Implementation of Policy Automated Provisioning Grid Foundation of Dynamic Collaboration

4.5 Technology Foundation quest comes from the client. The second application
The middleware was developed on top of Globus manages the business process. The application does
Tool Kit 3. The grid services are implemented as not require a network bandwidth.
OGSI services. For availability timing, we used Glo- Figures 7 and 8 show the physical and logical
bus Tool Kit 2 based MDS. network configuration of the demo system. The sys-
tem manages two VOs which have network devices
4.6 Concept Proof Demo Applications such as packet shaper, firewall and load balancer
To proof the provisioning grid concept, we imple- along with servers. Managing the template configura-
mented two applications on top of the grid middle- tion for the applications, the system allocates an ap-
ware. The first application is a video stream delivery propriate number of servers and bandwidths for the
application. It consumes a number of servers and applications.
bandwidth between the servers and client as a re-

Fig. 7 Physical configuration of the demo system.

Fig. 8 Logical network configuration of the demo system.

NEC Journal of Advanced Technology, Vol. 1, No. 1 61

01-a10.p65 Page 61 2004/03/15, 13:38 Adobe PageMaker 6.5J/Win

 Special Issue Advanced Technologies Driving Dynamic Collaboration Featuring System Technologies

5. CONCLUSION tion of the future use, ?????( )

We have successfully demonstrated that the de- 5) Deployment

sign concept of the provisioning grid works effectively Current implementation assumes that all the nec-
in a dynamically changing environment. However, essary software is appropriately installed prior to the
the current prototype is a limited implementation, usage. For large-scale resource allocation, the system
and many of the functionalities remain for future will have to support on the fly deployment as the
work. The system has been being enhanced toward resource is reserved or allocated. The deployment has
product level maturity. Mentioned below are some of to support application, middleware and OS layer. The
the future issues we are currently planning to solve. application deployment should include all the con-
tents data, database schema and other registry infor-
1) Resource Brokering Policy mation as well as application software. It also has to
In the current prototype, we implemented a mini- deploy infrastructure software such as OS, DBMS,
mum set of policies that is just enough to handle Application server and Web server.
simple resource brokering cases. Part of the weak-
ness comes with mediation of the resource request REFERENCES
from multiple VOs when the resources are exhausted
[1] I. Foster, D. Gannon and H. Kishimoto, The Open Grid
from the pool. The system requires a more sophisti-
Services Architecture, Global Grid Forum OGSA working
cated policy to judge the priority of the request from group, http://forge.gridforum.org/projects/ogsa-wg.
multiple VOs. [2] S. Tuecke, K. Czajkowski, et al., Open Grid Services
Infrastructure (OGSI) Version 1.0, Global Grid Forum
OGSI working group, http://forge.gridforum.org/projects/
2) Multiple ROs
ogsi-wg.
Current implementation supports a single RO. It is [3] R. Wolski, J. S. Plank, et al., Analyzing Market-based
critical to support multiple ROs for the future deploy- Resource Allocation Strategies for the Computational
ment of the large scale grid. This will require full Grid, The International Journal of High Performance
Computing Applications, 15, 3, pp.258-281, 2001.
implementation of the brokering model discussed in
[4] R. Buyya, H. Stockinger, et al., Economic Models for
this paper. Management of Resources in Peer-to-Peer and Grid Com-
puting, Proceedings of the SPIE International Confer-
3) Dynamic VO Lifecycle ence on Commercial Applications for High-Performance
Computing, Denver, 2001.
Current Implementation assumes static VOs. For
[5] G. Wasson and M. Humphrey, Towards Explicit Policy
more realistic environment of the dynamic collabora- Management in Virtual Organizations, Proceedings of
tion, VO has to be dynamically created and deleted as the IEEE 4th International Workshop on Policies for Dis-
required. tributed Systems and Networks, Lake Como, 2003.
[6] Utility data center: architecture, http://h30046.www3.
hp.com/solutions/architecture.html.
4) Resource Reservation
Current implementation assumes resource alloca-
tion for immediate use only. It does not support re-
source reservation. For supporting resource reserva- Received February 9, 2004

* * * * * * * * * * * * * * *

Takashi KOJO is now Chief Manager of Sys- Yoshiki SEO joined NEC Corporation in 19xx.
tem Platform Software Development Division. He is now Senior Manager of Internet Systems
He is one of the earliest members of Business Research Laboratories.
Grid Project jointly held by NEC, Fujitsu and
Hitachi. He is a co-chairi of CDDLM working
group in GGF. He has been in charge of many
of Web Services related development projects in NEC. He
received B.E. and M.E. of EECS, Aoyama-gakuin Univ.

Yoshiharu MAENO joined NEC Corporation in

19xx. He is now Assistant Manager of Internet
Systems Research Laboratories.

62 NEC J. of Adv. Tech., Jan. 2004

01-a10.p65 Page 62 2004/03/15, 13:38 Adobe PageMaker 6.5J/Win