Vikas Gorur wrote: > 2008/11/10 Harald St?rzebecher <haralds at cs.tu-berlin.de>: >> Beware that the network bandwidth might get divided between the >> connections to the servers. 8-) >> >> An extreme example: >> A client machine has one gigE connection to the network and uses >> client side AFR to replicate the data to 5 servers. > > This is a valid point. If you are concerned about the network being > the bottleneck, you can configure AFR on the server side and use a > separate gigE network for the replication. This is how we "solve" the bottleneck issue in our primary implementation of GlusterFS : two physical networks, one for normal traffic on eth0 (i.e. interaction with the web servers), and a second for storage traffic on eth1. Furthermore, we use server-side AFR, thus the clients never talk to each other (further reducing overhead). http://www.gluster.org/docs/index.php/High-availability_storage_using_server-side_AFR -- Daniel Maher <dma+gluster AT witbe DOT net>
