In best case distribute algorithm is O(1) and in case if there is a linkfile then it will be O(2) (which is treated as O(1) itself as per algorithm complexities). (and is not 'every' node as you mentioned), hence it won't be a big impact Regards, Amar On Fri, Mar 18, 2011 at 10:57 PM, Mohit Anchlia <mohitanchlia at gmail.com>wrote: > Thanks! Does linkfile concept cause any performance degradation since > now it will need to request every node if file exists or not? > > On Fri, Mar 18, 2011 at 10:19 AM, Amar Tumballi <amar at gluster.com> wrote: > > Hi Mohit, > > Sorry for the delay.. answer inline.. > >> > >> Thanks for getting this release out. I was going through the release > >> and didn't quite understand "fix layout by issuing re-balance" means. > >> Does it mean if server is added or removed and then new files created > >> in "existing" directory will not hash accross all the nodes (including > >> new nodes)? > >> > > > > Currently, the hashing range/layout is different per directory, and is > set > > during the mkdir(). Hence, when a new node is added, new files created > > under existing directory will not be hashed to new nodes. that is handled > by > > issuing a 'gluster rebalance' command. > > And when there is a 'remove-brick' command, the existing hash layout in > the > > directory will miss some ranges which belongs to those new nodes, hence > > again, you would need 'gluster rebalance' which does fix the layout > issues. > > > >> > >> And another question is it necessary to run "migrate-data" after > >> running "fix-layout"? If answer is "no" then after fixing layout > >> (fix-layout) wouldn't it confuse gluster hashing algorigthm in anyway > >> when accessing existing files? > >> > > > > answer to first part of the question is 'no' (ie, you don't need to run > > 'migrate-data' always after 'fix-layout'). Internally the hashing > algorithm > > is fool proof to handle it. We have a concept called linkfile which does > > this. > > Regards, > > Amar > > >
