On Fri, 2004-10-08 at 21:54 -0400, Daniel Phillips wrote: > I think the biggest part of this problem is just defining what's needed. > How hard could it be to implement? I think the simple methods of implementation are exhausted. At least, I'm running out. Here's the list of requirements so far: (1) The resource manager has to be able to operate without reading files during failover. This means no fork/exec, as this would constitute reading files (the csnap agent has prepared the resources for the csnap server already). Absolutely no scripts, as they're unpredictable as far as memory use is concerned. So, existing resource managers are basically out. (2) Ability to have randomly configured clusters with varying node performance, network topologies, and SAN/NAS topologies. No administrator intervention should be needed in the event of a failure in the non-uniformly-configured cluster case, as it could be inconvenient. This means the DLM model(s) are more or less out. (I'm sure there are more.) > Oh, is the resource manager to be distributed, or will it be a > server? ;-) Currently, it's "simple and dumb", but you can think of it as "distributed" if you like. It does not use group-decisions on where things should be placed. Currently, simple policies are used for failover which can be decided upon quickly by individual nodes without any input from other nodes (only the underlying infrastructure) at time of failover. -- Lon
