Tit-for-Tat Distributed Resource Allocation

Distributed computing infrastructures have risen in popularity over the past decade, however resource allocation on these—largely federated—systems remains a practically unsolved problem. Previous attempts have either relied upon central management of resources, which is infeasible in a federated model, and market-based economies which we classify into systems using global, transferrable currency and systems using local, transferrable currency. We present an alternate approach which represents pair-wise, tit-for-tat relationships as local, non-transferrable currency. This approach promises to provide a solution that does not require a globally trusted third party and is easier to deploy and maintain than previous systems. 1. PROBLEM DEFINITION Recently developed large-scale, distributed, federated computing infrastructures like PlanetLab [2], VINI [4] and the currently planned GENI [1] require resource allocation mechanisms which provide fairness, efficiency, and reasonable performance for each scheduled task while honoring federated control. In addressing this problem, we keep the following four goals in mind. First, a solution should work in a federated environment where there are many different administrative domains which may not entirely trust one another. Second, systems should not require a globally trusted third party as this explicitly undermines several of the attractive properties of federated environments. Third, per-domain configuration should be kept to a minimum and preferably eliminated as excessive configuration often leads to difficulties in deploying and maintaining systems. Fourth, the provided allocations should give reasonable performance with respect to some notions of fairness and efficiency. Previous work on distributed resource allocation can be coarsely categorized based on who issues credit or currency, and whether it can be moved. Systems like Bellagio [3] and Tycoon [7] use global, transferrable currency, which is issued by a globally trusted mint and easily transfered between parties. These systems cast resource allocation as a constrained optimization problem where each user has a utility function describing the locally perceived benefit of each possible resource allocation. With this formalism in place, it is easy to define efficiency as the ratio of the the total benefit of the current assignment to the maximum possible benefit and fairness as the minimum ratio between two different user’s benefits [5]. These two metrics are almost always at odds because the most efficient solution will give all of a given resource to whoever placed the highest value on it in their utility function. The global currency is used to solve this problem, by having each party bid on resources, thus constraining the set of possible solutions to ones which are both efficient and fair. The use of a globally trusted mint to produce all currency in effect gives the mint total power, which seems at odds with the goals of a federated infrastructure. SHARP [6], on the other hand does away with global currency by having each resource provider issue tickets which can later be probabilistically exchanged for resources. Furthermore, these tickets can be subdivided and reissued to other parties. In this way, SHARP uses local, transferrable currency and does not require global trust in a centralized mint. SHARP differs from Bellagio and Tycoon in that it provides a general mechanism for building such systems, but does not aim to explicitly define what policies should be deployed, instead providing some examples of possible configuration. Also, we believe that most of the complexity involved in SHARP comes from the fact that currency can be transferred requiring many mechanisms to secure the currency and agents which can reason about transferrable currency. While general properties of market-based systems and strategies can be analyzed in a game theoretic manner [5], they rely on accurate utility functions. Defining and maintaining such utility functions requires very smart algorithms, very smart people or both which violates our principle of minimal complexity and configuration. Even with an accurate local utility function, building a bidding-agent to maximize utility on a fixed budget is far from trivial. Because we feel that the minimal configuration and lack of global trust are extremely important for the adoption of a resource allocation scheme we propose an alternate strategy based on local, non-transferrable currency. The ability to transfer currency enables 3-way trades, which are critical in the situations where different entities have non-fungible offerings. On the other hand, in distributed computing infrastructures it is uncommon for the varying administrative domains to provide functionally distinguishable resources as resources usually vary in amount, but not in type. Instead, we maintain currency locally within each domain in the form of credit given to other domains for providing resources in the past, creating pair-wise, tit-for-tat relationships between administrative domains. 2. TIT-FOR-TAT ALLOCATION In this section, we propose a tit-for-tat strategy for exchanging resources. Each administrative domain has a set of one or more nodes which are most likely physical machines