[Gluster-devel] Improving real world performance by moving files closer to their target workloads

[Gordan Bobic]

Luke McGregor wrote:
>  - we would like to retain the previous copy also as long as there is
> free space on the system and have references to both files if this is
> possible. The idea would be that over time the nodes which use files
> regularly would all have copies of the particular file, (obviously
> there is a synchronisation problem here but this could be worked
> around), when space is needed the least read copy should be deleted
> (assuming that it isnt the last copy). Does this make sense, im not
> sure i have explained it very well.

It makes perfect sense. If this is always going to be in a LAN 
environment, you could probably get away with broadcasting a request for 
a file and seeing how many nodes respond. If you specify that you want 
redundancy of X (X copies of files in the network), you could then 
delete the least recently used local file when you are running out of 
space, provided that at least X other nodes have the file in their local 
stores.

In theory, as files migrate to local nodes, the number of broadcasts 
will reduce, as most requests will be satisfiable locally. The downside 
is that for each write, you'd have to find the peers with a copy of the 
file, and send them the delta to update locally. Locking may become 
difficult, and albeit doable, there may be some nasty race conditions to 
overcome.

Gordan