[1720] in linux-net channel archive
Re: Performance Enhancements to NFS client.
daemon@ATHENA.MIT.EDU (Olaf Kirch)
Sun Jan 28 10:26:32 1996
Date: Sat, 27 Jan 96 18:20 MET
From: okir@monad.swb.de (Olaf Kirch)
To: linux-net@vger.rutgers.edu
In article <199601260812.AAA24230@soda.CSUA.Berkeley.EDU> you wrote:
: Who else is working on the NFS client right now? What would be a good
: list on which to discuss NFS under linux?
I'm currently writing some stuff to do at least proper read caching with
the new page cache in 1.3.58.
: One of the things I might do (to decrease latency and possibly to do
: read ahead) is demultiplex incoming replies and send out
: (re)transmissions in the NET or TIMER bottom half interrupts (by
: overriding the NFS socket's sk->data_ready and by adding a timer).
: Can this be done? In particular, is there any reason why udp_sendmsg
: cannot be called during a bottom half interrupt?
These latencies are not the problem faced by the Linux NFS client right
now. It's the cache and readahead/writeback support that's missing.
OTOH, demultiplexing replies in data_ready may be worthwhile because it
avoids the kludgy select()/wakeup stuff and saves the udp_recvmsg(MSG_PEEK)
call.
Even if it worked, scheduling retransmits in the timer BH does by you
anything. When an RPC call times out after 0.6 seconds, a few
milliseconds extra won't be noticeable anyway.
Readahead does not belong in a bottom half, either, because you may have
to allocate additional pages, etc. Besides you'll be losing clock ticks
by the dozen.
Olaf
--
Olaf Kirch | --- o --- Nous sommes du soleil we love when we play
okir@monad.swb.de | / | \ sol.dhoop.naytheet.ah kin.ir.samse.qurax
