On the Scalability of Centralized Control

Scalability of clusters and MPPs is typically discussed in terms of limits on growth: something which grows at a rate of O(log p) (where p is the number of processors) is said to be more scalable than something whose growth rate is O(p). But in practice p does not grow without limits. We therefore suggest that discussions of scalability should take time into account. System sizes grow with time, so larger systems need to be supported — but only after some time. And in particular, there is no real need to support arbitrarily large systems right now. Surprisingly, when time is thus put into the picture, we find that centralized control is actually quite scalable. The reason is that the capabilities of a centralized control node grow at a fast pace due to Moore’s law. This seems to be more than enough in order to manage current growth patterns displayed by parallel systems.

• Type: Conference
• Year: 2005
• Where: IPPS
• Authors: Dror G. Feitelson
• Keywords: Centralized Control, Distributed And Parallel Computing, Growth Rate, IPPS 2005, Scalability