On 12/17/2011 9:41 PM, Garance A Drosihn wrote: > Hello. A minor question here. > > I was curious if there was some way to encourage clients to load- > balance their references across multiple RO volumes. I have a > volume which is fairly large (for us), and is often referenced > about 10,000 times an hour. I wanted to move that volume to new > fileserver partitions. > >[...] > > It's probably true that this volume is referenced from a small > number of client machines (web servers), and those machines are > rarely rebooted. But is there some way to encourage the clients > to balance out the references? The volume hasn't been very busy > today, but I've seen other days where the RO on afsfs13/a has been > referenced 500,000 times, and the one on afsfs12/b less than 2,000 > times. > > I do plan to move these specific volumes around some more over the > winter break, so the state of this volume will be changing anyway. > But it seemed like an interesting question.
The behavior that is getting the Unix clients into this situation is the vos release. At the point the release occurs the RO on the same partition as the RW becomes the only up to date version of the volume. Any clients that attempt to read from the RO during the incremental dump to the other RO sites will be forced to read the data from the source machine. Once a client is reading from a server it stays on that server. (The windows client behaves differently and will only stay on the server it is reading from until the volume location data is refreshed after two hours.) At the present time there is no method of avoiding this side effect of a volume release. However, Derrick (and Jake over the Summer) have implemented a new vos release option "-stayup" which ensures that all of the replication sites are updated simultaneously. See gerrit patchsets 6249 through 6256. This code is still being tested but you are welcome to take a look. Jeffrey Altman
signature.asc
Description: OpenPGP digital signature
