Proportion - share Resource

Proportional-share scheduling has been investigated by many researchers including: Wald-spurger et. al. at MIT and by Stoica et. al. at Old Dominion University. This work, however, suuers from high overhead and limited scalability problems when applied to a distributed system. This paper presents an alternative approach to proportion-share scheduling, which overcomes these problems. Our system provides a `marketplace' of resources formed from many local 'markets'. Each process is allocated a wage (dependent on importance) which they spend on resources they need, by transferring money to servers. To overcome the need for the bottlenecks caused by auctions or lotteries used in previous systems, each server instead directly allocates clients their potential share of access dependent on the total paid and current cost price of the resource. These cost prices are then periodically advertised by the servers using agent processes, which update other local markets in their locality. Thus each local market dynamically contains advertising from those important enough or close enough to be able to aaord to advertise there. Thus scalability is maintained whilst increasing the information spread over traditional nearest-neighbour meshes. Results of a prototype simulation of a distributed ERA system are presented, showing a maximum overhead of the ERA system of around 3.5 percent of execution time as the frequency of wage updates and number of processes are varied. Whilst scalability of the distributed marketplace remains linear under heavy separate workloads as the numbers of nodes in the system is varied.