Final Activity Report Summary - COMIG (The Content Migration (CoMig) paradigm for internet content dissemination)
We posed and answered the following two questions:
1. how could we design a 'content migration' paradigm for allowing digitised content to migrate to the most appropriate locations in a content dissemination network without requiring intervention from a centralised authority; and
2. taking for granted that the content dissemination network was run by multiple independent authorities, each catering to optimise its own local utility, what were the content replication and caching schemes that would arise and what was their relationship to traditional, 'socially optimal' schemes.
To address the first question, we casted the dissemination of content to a community of users as a facility location problem and developed new distributed solutions to overcome problems of previous solutions that were centralised and, therefore, impossible to apply in large networks with thousands of users and millions of distinct information objects. We used a 'divide and conquer' approach in which one started with an initial placement of copies which was then refined iteratively through local re-optimisations. We showed that the resulting final placement of content performed nearly as well as one produced by a traditional centralised approach that did not scale as well as ours.
Addressing the second question involved two parts. Firstly, we tackled the case of replication, in which permanent copies of objects were kept in the nodes of the network. We modelled the contention between different users as a strategic game and derived stable object replication schemes in the sense of a pure Nash equilibrium. We also showed that such schemes were realisable in practice in a distributed manner that required the exchange of a modest amount of information. We then addressed the same problem under object caching, in which case a copy of an object was stored only after a user had requested so. Caching also implied the existence of a replacement algorithm for freeing space when new objects were requested and the nodes were full. We designed a caching protocol that made the nodes cache objects in a way that was beneficial to their local utility. Thereby we approximated in an on-line fashion the stable object replication schemes, without, however, having to know demand and topology information that was necessary in the case of replication but rather hard to maintain when the characteristics of the network changed rapidly.
1. how could we design a 'content migration' paradigm for allowing digitised content to migrate to the most appropriate locations in a content dissemination network without requiring intervention from a centralised authority; and
2. taking for granted that the content dissemination network was run by multiple independent authorities, each catering to optimise its own local utility, what were the content replication and caching schemes that would arise and what was their relationship to traditional, 'socially optimal' schemes.
To address the first question, we casted the dissemination of content to a community of users as a facility location problem and developed new distributed solutions to overcome problems of previous solutions that were centralised and, therefore, impossible to apply in large networks with thousands of users and millions of distinct information objects. We used a 'divide and conquer' approach in which one started with an initial placement of copies which was then refined iteratively through local re-optimisations. We showed that the resulting final placement of content performed nearly as well as one produced by a traditional centralised approach that did not scale as well as ours.
Addressing the second question involved two parts. Firstly, we tackled the case of replication, in which permanent copies of objects were kept in the nodes of the network. We modelled the contention between different users as a strategic game and derived stable object replication schemes in the sense of a pure Nash equilibrium. We also showed that such schemes were realisable in practice in a distributed manner that required the exchange of a modest amount of information. We then addressed the same problem under object caching, in which case a copy of an object was stored only after a user had requested so. Caching also implied the existence of a replacement algorithm for freeing space when new objects were requested and the nodes were full. We designed a caching protocol that made the nodes cache objects in a way that was beneficial to their local utility. Thereby we approximated in an on-line fashion the stable object replication schemes, without, however, having to know demand and topology information that was necessary in the case of replication but rather hard to maintain when the characteristics of the network changed rapidly.