Just in case anyone is interested, here's a little background on
write gathering (see Chet's paper for the whole skinny).

For V2 clients with lots of biods/nfsiods running, the clients would
generate bursts of write RPCs close to-gether.
- since a server must commit the data and metadata changes to stable
storage before replying to the write RPC for V2
--> waiting a little while for more write RPCs and then doing then
all to-gether and replying to all of them resulted in less server
disk I/O under heavy load

As such, a server with a heavy load of V2 clients could handle a larger
write load (disks were much slower 10 years ago and could be max'd out
easily).

As an alternative to write gathering, several companies sold
non-volatile ram stable storage devices to improve server write perf.

The V3 protocol fixed the problem, by allowing write RPCs to be done
asynchronously on the server (no requirement for write data and metadata
to be committed to stable storage), then a commit RPC followed later.
For this case, write gathering provides no benefit and almost guarantees
slower response to client write RPCs. (btw, some vendors used to cheat
and not actually commit the data and metadata to stable storage, but
that did violate the V2 RFC.)

So, now that almost everything is V3, I really believe that write
gathering should go the way of the dodo. Oh, and V4 uses the same
write mechanisms as V3, so it has no use for write gathering, either.

Have fun with it, rick