Multiple Distributed Auctions for Allocating Grid Resources Peter - - PowerPoint PPT Presentation

Multiple Distributed Auctions for Allocating Grid Resources

Peter Gradwell and Julian Padget Department of Computer Science, University of Bath, Bath, UK

Multiple Distributed Auctionsfor Allocating Grid Resources – p.1/6

Market-based Resource Allocation

Trading systems are no use if they are slower Need for an accurate empirical model of Combinatorial Auctions: ⊲ Algorithms CABOB, CASS, LP Solve ⊲ NP-Complete ⊲ Hard increasing the number of goods makes the computation time much longer than increasing the number of bids (Fujisima, Leyton-Brown, Shoham) Aid understanding of when to use different market mechanisms (CAs, CDA, Distributed Markets etc.)

Multiple Distributed Auctionsfor Allocating Grid Resources – p.2/6

Parameters and Complexity

20 40 60 80 100 120 140 160 50 100 150 200 250 300 350 50 100 150 200 250 Time (ms) Plot of Goods, Bids, Time Bids = 2xGoods Bids = Goods Goods Bids Time (ms)

Number of goods has more impact on computation than number of bidders. Literature demonstrates that some problems respond to heuristics, but others do not Criteria for choosing market or auction still unclear

Multiple Distributed Auctionsfor Allocating Grid Resources – p.3/6

Distributed Auctions

A market-based solution Distributed Auctions enable cross-fertilisation of a wide range of traders and buyers — as found in grids. Intelligent (middle) agents assemble bundles against customer requirements (actual or prospective) Trader agents are profit motivated. Traders may not sell all their bundles — so there is natural wastage in the system. Multiple Distributed Auctions (MDAs) are suitable for

• pen grids as no relationship is required between

trading parties

Multiple Distributed Auctionsfor Allocating Grid Resources – p.4/6

CAs vs Distributed Systems

Time Bundle Complexity Perfect Allocation CA M D A

Complexity can neither be created nor destroyed If we remove the single combinatorial auction, who does the computation? Intelligent (middle) agents assemble bundles against customer requirements (actual or prospective) ⊲ Cost is distributed ⊲ Optimality is forfeited ⊲ Worse is better?

Multiple Distributed Auctionsfor Allocating Grid Resources – p.5/6

Approaching Optimality

Current work: investigating the proximity of a MDA bundle to the (strongly) Pareto-optimal bundle. The depth of search (and speed of result) obtainable by a CA clearing algorithms are highly dependent on the heuristics used in the computation (CABOB). The MDA approach is very unlikely to produce a Pareto-optimal solution because has it has incomplete information Can heuristics be used to improve the MDA bundling mechanism? Could MDA traders remember popular bundles and assemble them pre-emptively? Market memory. How does re-sale/re-circulation of items impact market dynamics?

Multiple Distributed Auctionsfor Allocating Grid Resources – p.6/6