On Fri, 6 Feb 2009 15:57:57 +0530, Vikas Gorur <vikas@xxxxxxxxxxxxx> wrote: > Gordon, > > (Replying to both your mails) > > Implementing multicast would be difficult since multicasting only > works for UDP, and GlusterFS relies on a TCP connection. Moving GlusterFS > to use UDP would require extensive changes and would bring many new > problems of its own. > > Regarding the configuration you'd want to conserve bandwidth for some of > your > machines, you could do something like this (shown for one machine): > > simple nfs-like > glusterfs afr > low b/w box --------> /mnt/glusterfs --------> other three boxen > > high b/w box > > I don't see how this helps, though. The high b/w box has to send writes > to the other three machines anyway, thus robbing them of their bandwidth? The main point to this would be to ensure that each machine doesn't have to replicate ALL it's writes to ALL peers - instead, it could replicate them to _one_ central high-bandwidth peer which then replicates to other satellite peers. This means that the bandwidth is not limited to 1/n of upstream bandwdith of satellite peers - instead the full upstream bandwidth would be achieved. True, the central peer still consumes n times more bandwidth (replicates full bandwidth to each satellite peer), but the main point is to overcome the satellite peer bottleneck by avoiding the write bandwidth multiplication that causes the scalability collapse. Gordan
