From: Stas O. <sta...@gm...> - 2010-08-16 09:04:49
|
Hi. hmmmm. master decides where data are to be written. Chunkservers read > or write data when they are commanded. Chunkservers do not, afaik, take > decisions on their own about read or write. > The idea is in case the application running on chunk master, it will write and read locally for improved performance. I think you're doing a mix-up between rack and subnet. you can > perfectly have the same subnet in several racks. > Anyway, rack awareness is not yet implemented. It's on the todo list as > far as I can remember. (can't verify as I can't access web for now). > You right - this specified on the chunk server level as additional ID. The idea is to have reads in same rack, but replications across racks. This increase performance and lowers Intra/Inter net bandwidth usage, while still providing required level of fault tolerance. These two features are being implemented in many DFS systems these days, and I think they will be a great addition to MFS as well. For example: http://kosmosfs.sourceforge.net/features.html " Rack-aware data placement: The chunk placement algorithm is rack-aware. Whereever possible, it places chunks on different racks. . . Local read optimization: When applications are run on the same nodes as chunkservers, the CloudStore client library contains an optimization for reading data locally. That is, if the chunk is stored on the same node as the one on which the application is executing, data is read from the local node. " Regards. |