Optimizing Decentralized Grid Markets through Group Selection Isaac Chaoa, Oscar Ardaizb, Ramon Sangüesaa c

Liviu Joita , Omer F. Rana a

c

Computer Architecture Department, Polytechnic University of Catalonia, Spain {ichao,sanguesa}@lsi.upc.edu

b

Department of Mathematics and Informatics, Public University of Navarra, Spain [email protected]

c

School of Computer Science and the Welsh eScience Centre, Cardiff University, UK {l.joita, o.f.rana} @cs.cardiff.ac.uk

Abstract Automatic coordination mechanisms for the Grid are required due to the increasing complexity exhibited in large scale distributed systems. Decentralized economic models are being considered as scalable coordination mechanisms for the management of service allocations to clients. However, decentralization incorporates further dynamicity and unpredictability into the system. Introducing higher levels of adaptation and learning in the coordination protocols helps cope with complexity. We provide a solution based on a self-organized, emergent mechanism evolving Grid Market participants through a group selection process. Dynamic congregations organize agents into optimized market segments, maximizing utility and thereby improving system-wide performance. We provide a system model and evaluation by simulation of the group selection mechanism. We further provide a prototype showing practical feasibility of the approach.

1. Introduction Virtual Organizations (VOs) have the potential to change dramatically the way we use computers to solve problems, much as the Web has changed how we exchange information [1]. One particular use of such VOs has been to coordinate resource usage within Grid environments, enabling controlled interaction between users and resource providers distributed across multiple administrative domains. Decentralized Grid markets utilizing multiagent system techniques have been proposed as approaches for coordination within such VOs [2]. The market here is nothing more than a communication bus and does not participate in matching participant’s requirements. Direct agent to agent bargaining allows participants to use a negotiation strategy that is more suitable, based on their objectives and current circumstances. Local bilateral bargaining also facilitates the scalability of the system and enables adaptation to fluctuations in resource allocation dynamics. However, Grid markets still lack some responsiveness to a user’s perceived utility. For instance, suppose we were to build a spot Grid

market with a given initial configuration in order to better allocate a number of services to users. The market evolves over time and bilateral negotiations take place between service provider agents and user agents or brokers acting on behalf of users, eventually achieving an efficient allocation of resources. However, after services get executed and effective utilities are derived, how can we react in order to improve these utilities? In such a scenario, agents could try modifying prices/budgets and negotiate again in the same market using a different strategy, but it is also possible to consider how such agents would perform in another market or with a different set of trade partners. Bootstrapping in market segments and further co-evolution of the Grid market segments is given little attention in state-of-the-art Grid economics research. Group Selection refers to a process of natural selection that favors characteristics in individuals that increase the fitness of the group. Therefore, Group Selection implies that every member of the group depends on a group characteristic that is not isolated in a single individual. We provide a coordination mechanism based on Group Selection which can be used

standalone or alongside an existing coordination architecture (such as a market) in order to optimize its performance. The key idea is that biasing interaction between Grid nodes by arbitrary “Tags” or identifiers enables efficient grouping of agents into appropriate market segments. Further evolution through Group Selection optimizes groups in performance. Moreover, such segmentation has many desirable properties: it is adaptive, self-organized, decentralized and highly scalable. In addition, the approach can be used to automatically manage different stages (formation, evolution and disbanding) of a VO lifecycle. The rest of the paper is organized as follows. In section 2 we review related work. In section 3 the Group Selection model for optimizing Grid markets is provided, along with a proof-of concept prototype. In section 4 we provide experimental results and evaluation of the approach. Section 5 concludes the paper.

2. Related Work Economics-based resource allocation has received a great deal of attention in recent years. The GridBus Project [3] is a reference in Grid Economics and utility-based computing, and has proposed a great variety of market models and tools for the trading of Grid Resources. However, its strong emphasis on computation-intensive Grids and the hierarchical nature of some of the proposed components, like the Grid Market Directory, diverges from the fully decentralized resource allocation mechanisms proposed here. An alternative, fully decentralized, approach is the one adopted by the Catallaxy-based agents [2]. In this approach bilateral negotiations are established between a set of learning agents, and the spontaneous coordination arises from both the bargaining and co-evolutionary learning processes. The Catallaxy mechanism was originally proposed by von Hayek [4], as a coordination mechanism for systems consisting of autonomous decentralized agents that makes use of a “free-market” approach. It enables prices within the market to be adjusted based on constant negotiation and price signaling between agents. Catallactic agents for Grid resource allocation are being explored in the CATNETS project [5]. The model proposed in this paper attempts to extend standard Catallaxy-based Grid markets, enabling the bootstrapping of the agents in sub-markets and the automatic evolution of these groups towards optimized market segments. Hayek also supported the idea of

Group Selection as a transmitter of free market norms and institutions between societies [6]. Exploiting group structure in multi-agent systems (MAS) has also been proposed in previous research. Most of the coalition formation literature attempts to formalize optimal grouping mechanisms for agent populations. Major limitations of these algorithms are a high computational complexity, and unrealistic assumptions regarding the availability of information [7]. Not being super-additive (the value of unified coalitions is not necessarily greater that the sum of the two composing coalitions), markets are a very demanding scenario for state-of-the-art coalition algorithms. These issues limit the practical use of these mechanisms in large scale scenarios such as Grids. An alternative approach is that of forming “congregations” [8], which can be used to serve as market optimizers. However, congregations are static and agents can only trade in a specified number of sub-groups. In contrast, group selection approaches enable a dynamic view of the system, evolving the required number of subgroups depending on the changing agent’s needs.

3. Optimizing Decentralized Grid Markets trough Group Selection Recent trends in both distributed systems and MAS try to tackle the problem of coordination. In MAS this is achieved from a bottom-up point of view, by incorporating loose coupling between processes/agents and emergence of the solution to the coordination problem. This approach proves to be relatively tolerant to failures in individual processes/agents. Group Selection complies with the characteristics above for emergent coordination mechanisms. In evolutionary biology, Group Selection refers to the idea that “alleles” can become fixed or spread in a population because of the benefits they bestow on groups, regardless of the fitness of individuals within that group. The theory of Group Selection states that selective forces can in fact act on competing groups of individuals, not just competing individuals [9]. It has been shown that Group Selection can lead to the spread of mutually beneficial characteristics in agent populations within a multi-group setting [10]. This enables the application of Group Selection processes in any group-like structured agent population. In engineering, novel socially-inspired mechanisms have been developed, building on Group Selection processes. These mechanisms use a Tag (or social label) to identify groups.

Agent interactions are biased by Tags (i.e. within the groups) and inter-group migrations are ruled by Group Selection processes. Notably, the use of a Tag-based mechanism has been demonstrated to reduce the free-riding problem in P2P networks [11]. Also other applications build on Tag-based mechanisms such as a mixed Tag/sanctioning mechanism for free-riding prevention in VO coordination [12]. It is intuitively clear that decentralized bidding in a flat market is not the most efficient setting for any given number of buyers and sellers in a Grid Market. The higher the heterogeneity and dynamicity of the system, the bigger is the probability of un-coordinated outcomes leading to low system-wide performance. The solution proposed here is a

mechanism building on Tag identifiers to segment a Grid market in a self-organized manner. The idea behind market segmentation is to identify groups of similar customers, to prioritize the groups to address and to respond to appropriate strategies that satisfy the different preferences of each segment. Once the negotiation partner is chosen, the negotiation proceeds in exactly the same way as it would do in a normal Catallaxy-based market. Market segments could be set up manually, but this clearly does not scale nor adapt to dynamically changing VOs. The key benefit is the fact that Group Selection mechanisms work with large scale open systems, such as the systems targeted in Catallaxy-based markets.

Grid Market

Grid Market Segment1

Service1

negotiation

strategy S1

Service1

strategy S1

negotiation

Client1 negotiation

Client1

Service2

S1?

S1? negotiation

negotiation

Service3

strategy S1 Service2

Migrate?

strategy S1

Segment2 Service3

strategy S2

negotiation

Client2

Client2 negotiation

S2?

negotiation

strategy S2

Service4

negotiation

strategy S2

S2?

Service4

strategy S2

Figure 1: (left)Flat Catallactic Grid Market; (right) Segmented Grid Market A minimal segmentation of the Grid market is depicted in Figure 1. Manual configuration in this context would involve grouping complementary agents together. The power of the Group Selection mechanism lies in its ability to scale with the number of agents and changes in market structure. After one market iteration, and after utilities from service executions have been calculated, a user sends back his perceived utility. This can be a simple metric such as the service provision time, to more elaborate metrics comprising other quality attributes. Then Clients and Services must decide how to evolve on the

different market segments. From the initial configuration they can evolve by comparing fitness with agents in other VOs, or migrating to VOs where they find outperforming agents. In this scenario, it is this closeness in “negotiation type” which is used as a basis to automatically select suitable partners at the group level. In this manner Client 1 and Client 2 will interact preferentially with the agents sharing the same Tag (i.e. belonging to the same group), which tends to be agents closer to them in negotiation abilities and goals, hence increasing the probability of successful allocations.

Bootstrap agents in groups LOOP a number of rounds

DataSet Migrate

Convert

Converter Service B Instance

Data-Mining Execute

LOOP each group

Mining Service A Instance

Results Output

Resource Allocation

LOOP each agent in the group (operation phase) Apply Interaction rule: Negotiate with agent/agents from the submarket Collect Payoffs from allocations ENDLOOP ENDLOOP

Converter Service A

Mining Service A

Resource Provider

CS Agent1

MS Agent1

R Agent1

Converter Service B

Mining Service B

CS Agent2

MS Agent2

Resource Provider R Agent2

Converter Service C

Mining Service C

CS Agent3

MS Agent3

(MGS) Complex Service Agent

LOOP each agent in the population (evolution phase) Select random agent in the entire population Apply Migration rule:

Tag 1 (VO1) Negotiation messages

If(randomAgent outperforms me)

Tag 2 (VO2)

then Copy its negotiation type and move to its group

Resource Provider R Agent3

Negotiation messages

ENDLOOP Service Market

ENDLOOP

Resource Market

Figure 2: (a) Group Selection Mechanism; (b) Application to Data-Mining prototype The current synchronous realization (figure 2, left) does not prevent for applications in a realistic, asynchronous environment, since no explicit synchronization points are required. The bilateral bargaining proceeds in accordance with the Catallaxy-based approach during the interaction phase. The only change is the scope of interaction, now being the submarket instead of the whole population. A key theme is the evolution to other groups through migration and the copy of the negotiation type of fittest in the evolution phase. This is what spreads compatible negotiation types among the market segments. This process is selforganized and scales automatically to any given market size, segmenting the flat market into a number of groups. As a proof-of-concept of the system model, we have adapted a data mining application to demonstrate the Tag mechanism [13]. The data mining process is often structured into a discovery pipeline/workflow, involving access, integration and analysis of data from disparate sources. In the data mining prototype, two services “data conversion” and “algorithm execution” are combined into a workflow. Consider a scenario (figure 2, right) where a Client (Master Grid Service in the Catallaxybased market jargon) issues many sequential requests to data mining services. Brokers (Complex Services or CSs in the Catallaxybased market jargon) try to map incoming workflows to a set of available services. Service Providers (BSs in the Catallaxy-based market jargon) try to sell their service to the

CSs. The greater the heterogeneity in the service requirements is, the harder the probability that a given decentralized search of trade partners finds compatible pairings in the flat market. In the simplest scenario, two Tags are shown in different circles. These represent two disjoint negotiation spaces (agents with Tag1 interact preferentially with agents possessing the same tag, and the same for Tag2). After multiple runs, the CS allocates successive workflows which derive higher utility for the clients. The evolution of the VOs executing the workflow is driven by the Group Selection process. For a large number of Tags (hence for a big number of services and VOs), compatible CSs and BSs are expected to be grouped into the same workflow, hence optimizing the service selection provided by the flat Catallaxy-based market. Such a feature would be hard (if not impossible) to achieve in a manual manner.

4. Experimental evaluation

results

and

We have used an implementation of the ContractNet protocol, standardized by FIPA, [14], to support decentralized economic agents. An example of the use of Contract-Net for resource allocation in VOs can be found in [15]. The Contract-Net protocol starts with a task announcement phase by the initiator (the buyer), which can be answered by one or more participants (the sellers). This announcement is carried out by the “groupcast” (equivalent to a multicast) of a call for proposals (CFP). After conclusion of this first period, the initiator selects from the set of collected

proposals the best one, informing the winner. On top of this protocol, we apply a simple offer/demand-based economic algorithm. The sellers will answer the CFPs which meet its current selling price. If the CFP does not meet its requirements, the seller will lower its expectations and will decrease the selling price. As for the buyers, if a seller rejects the CFP, then it will lower its expectation by increasing the offer in the next CFP. Both the buyers and the sellers will increase their expectations in case of receiving offers/bids which meet their expectations. The price update process is done at fixed small price steps. Initial prices for the agents are set to random values in the bounds 73 and 77, for a mean initial price of 75. This range is arbitrarily set. The experiments are conducted on a generic agent-based Grid simulator, developed Accumulated Traders Utility

for modeling agent coordination mechanism within Grid systems [16]. The experiments conducted here are repeatable by downloading the simulator. We set up a scenario with 100 agents (50 buyers and 50 sellers) which trade to achieve service allocations. We introduce a fixed rate of incoming requests, 1 request per simulation round issued per CS. Each request will trigger a CFP sent by each CS. In order to model realistic scenarios with communication costs, we limit the scope of the CFP to 5 BSs, to be reachable on each groupcast, selected at random either from the whole population (in the flat scenario) or from agents belonging to the same group (in the group selection scenario). We introduce five different “types of negotiation”, modeling the different contractual implications, legal issues and so on of the traded services. Utility derived by clients is related to the “closeness” between negotiation types. The closer the types, the higher the utility derived.

Segmented market

Average Market Prices Group Selection

Flat market

Buyers Sellers

70 60

40

P ric e

Utility

50

30 20 10 0 0

20

40

60

80

100

time (simulation tick x10-1)

80 79 78 77 76 75 74 73 72 71 70 0

20

40

60

80

100

time (simulation tick x10-1)

3: (left) Using Group Selection, allocation utility for traders is optimized. (right) Prices remain stable in both the flat and segmented market. The x-axis in both graphs reports simulation time. We measure the utility extracted from allocations as directly proportional to the negotiation type closeness.

U=

1 nt1 − nt 2

Given M different negotiation types, negotiation types are represented by integers from 1 to M. We use (2) for utility calculation, with nt1 being the negotiation type of buyer

nt agent 1 and 2 being the negotiation type of seller agent 2. The exception is when nt1 = nt 2 , in which case we consider the

maximum utility of 1 having been achieved. The rest of the cases oscillate between 0 and 1. If we compare a flat market with a segmented market evolving through Group Selection, we see how the latter improve coordination, leading to better accumulated allocation utilities (figure 3, left). Prices vary smoothly influenced by offer/demand in the bounds 70 to 80 in the Group Selection scenario. The price stability around the initial selling prices (75) renders the market fair (figure 3, right). The price evolution for the baseline flat market is similar; hence we do not show the graph. We can conclude that market segmentation is able to increase resource allocation utilities of the traders without compromising price stability.

7. Conclusions In large scale VOs, system dynamicity and uncertainty are high; therefore automatic, decentralized and self-organized control becomes a requirement. Our proposal is that a simple, rather powerful coordination mechanism based on Group Selection can be used to self-organize a set of agents in VOs to operate more effectively within a Grid environment. The mechanism complements rather than competes with much of the existent coordination mechanisms for multi-agent systems. In the case of plugging the mechanism into an economic Grid middleware, each VO represents a market segment, and migration of agents from one sub-market to another is ruled by the Group Selection process, achieving optimization of allocation utilities. We provide a proof-ofconcept prototype application demonstrating the model in a real decentralized market prototype, trading data mining services. The results from simulations on an agentbased Grid simulator are higher profits for the society of agents trading in decentralized markets which structure the population in submarkets and incorporate Group Selection, compared with learning agents trading in flat markets. Both mechanisms do not impose any constraint on system size or computational requirements, hence enabling for high scalability in both physical and organizational dimensions. Future work includes considering users and services belonging to many VOs simultaneously, hence trading in different market segments at the same time. Also, extending the economic algorithms to include other more elaborate bargaining strategies from the CATNETS project [5] can help generalizing the results of this paper.

8. References [1] Ian Foster, Nicholas R. Jennings and Carl Kesselman, "Brain Meets Brawn: Why Grid and Agents Need Each Other" in Proceedings of 3rd Int. Conf. on Autonomous Agents and Multi-Agent Systems (AAMAS 2004), New York, USA, [2] T. Eymann, B. Padovan, and D. Schoder. The catallaxy asa new paradigm for the design of information systems. In16th IFIP World Computer Congress, Beijing, China, August2000. [3] Buyya, R. Abramson, D. Venugopal, S, The Grid Economy, Proceedings of the IEEE pages 698714 Volume: 93, Issue: 3 ISSN: 0018-9219, 2005

[4] Hayek F. A., Bartley W., Klein P., Caldwell B., The collected works of F. A. Hayek. University of Chicago Press, 1989 [5] CATNETS Project IST-FP6-003769 (2006), http://www.catnets.org [6] Zywicki, Todd J., "Was Hayek Right About Group Selection After All? Review Essay of Unto Others: The Evolution and Psychology of Unselfish Behavior" (1999). Review of Austrian Economics Available at SSRN: http://ssrn.com/abstract=182509 or DOI: 10.2139/ssrn.182509 [7] Coalition Formation: Towards Feasible Solutions. Fundamenta Informaticae, Vol. 63, No. 2 - 3. (January 2004), pp. 107-124 [8] Chistopher H. Brooks and Edmund H. Durfee. Congregating and market formation. In Proceedings of the 1st International Joint Conference on Autonomous Agents and MultiAgent Systems, pages 96-103, 2002. [9] Wilson, D.S. (1975). A theory of Group Selection. Proc. Nat. Acad. Sci. USA 72: 143-146. [10] Boyd, R. and Richerson, P. (2002) “Group Beneficial Norms Can Spread Rapidly in a Structured Population.” Journal of Theoretical Biology 215. 287– 296 [11] Hales, D. (2004) From Selfish Nodes to Cooperative Networks – Emergent Link-based Incentives in Peer-toPeer Networks. In proceedings of The Fourth IEEE International Conference on Peer-to-Peer Computing (p2p2004), 25-27 August 2004, Zurich, Switzerland. IEEE Computer Society Press [12] Isaac Chao, Oscar Ardaiz, Ramon Sanguesa . Tag Mechanisms Applied to Open Grid Virtual Organizations Management, in Proceedings of the Joint Smart Grid Technologies (SGT) and Engineering Emergence for Autonomic Systems (EEAS) Workshop, Editors: R. Anthony, A. Butler, M. Ibrahim, T. Eymann, and D. J. Veit, Dublin, Ireland, pp 22-29, 2006 [13] Liviu Joita, Omer F. Rana, Pablo Chacin, Isaac Chao, Felix Freitag, Leandro Navarro, Oscar Ardaiz (2006) - "A Catallactic Market for Data Mining Services", in International Journal of Future Generation Computer Systems (FGCS) - Grid Computing:Theory, Methods & Applications, Volume 23, Issue 1, January 2007, ISSN 0167-739x, pp. 146-153 [14] FIPA webpage: http://www.fipa.org/ [15] Isaac Chao, Ramon Sangüesa and Oscar Ardaiz, Design, Implementation and Evaluation of a Resource Management Multiagent.System for a Multimedia Processing Grid, Workshop on Grid Computing and Its Application to Data Analysis (GADA) On the Move to Meaningful Internet Systems 2004. [16] https://sourceforge.net/projects/agentgridrepast