[ale] networking backpressure

Derek Atkins warlord at MIT.EDU
Wed Apr 10 11:05:16 EDT 2013


Jim,

Jim Kinney <jim.kinney at gmail.com> writes:

> so what happens when a series of machines, each with 1G nic all try to read/
> write to the same NFS mount filesystem, also on a single 1G nic, and this goes
> on for hours at a time?
>
> Eventually, nic buffers fill and the OS is told to "hold that thought". But
> for how long? Will the OS continue to try the write until it succeeds? Or will
> the constant buffer overload just cause a nic to flush the buffer and hope for
> better luck next send?

Are you using UDP or TCP?

If you're doing TCP then you just have to worry about buffer bloat.
Beyond that, TCP will handle congestion control and should effectively
rate-limit all clients.

Using UDP, however, it's all a question of NFS's built-in congestion
control.  This would depend on the version of the client and server.

If your server is Windows, then I think you're probably screwed.  Their
NFS server implementation has never been good.

-derek
-- 
       Derek Atkins, SB '93 MIT EE, SM '95 MIT Media Laboratory
       Member, MIT Student Information Processing Board  (SIPB)
       URL: http://web.mit.edu/warlord/    PP-ASEL-IA     N1NWH
       warlord at MIT.EDU                        PGP key available


More information about the Ale mailing list