Re: [opensuse] OpenSuse 10.2 - Fortran compilation very slow through NFS network with a 64bit server and 32bit clients.

M Harris wrote:
On Wednesday 19 December 2007 09:44, Dave Howorth wrote:
but
ordering of packets over the network isn't a problem. And yes, either
TCP or UDP can be used.
Thanks. You can tell how long its been since I even tried it... back not many years ago only udp worked. tcp will keep the packets ordered. And thanks for the link.
----
UDP is stateless (no connection needed) -- but it is a reliable
transport for NFS since NFS keeps track of the packet ordering and the
packets which are 'outstanding'.

TCP is preferred and is the default when it is available.

Where TCP gets important is when NFS run between two
machines on a *high* speed network (1G or greater). The reason being
is that NFS uses a 16-bit counter to order packets. On a high
speed network, it's possible too many packets (>65535) can be sent'
out in a short time and NFS can get 'confused' and lose a bunch of
packets. This is why "tcp" is the default and preferred transport. TCP
is available with versions 3 & 4 of NFS (both supported by most modern
kernels). A big change in 10.3 (I think), is that 10.3 appears to
prefer NFS4 if available over NFS3. NFS "sessions" will use the maximum
supported.

However, regarding this:
Perhaps things would improve using "async", that's a difference with local filesystems that run "async" by default. But the man page says
it can be dangerous.

VERY!
---
It depends on the environment, but in a local area network,
"async" is usually preferred. "Async" causes *great* speed increases
when used with NFS that cannot be attained with a 'synchronous'
mount. Few 'end users' need 'sync' on NFS mounts.



--
To unsubscribe, e-mail: opensuse+unsubscribe@xxxxxxxxxxxx
For additional commands, e-mail: opensuse+help@xxxxxxxxxxxx

Relevant Pages

  • 2.6.0-test7-9 NFS performance problems
    ... I have verified that the problem isn't with the network card by using ... packets that it failed to reassemble, ... Server nfs v2: ... # ACPI Support ...
    (Linux-Kernel)
  • Re: Socket send blocking problem
    ... The network becomes congested, ... sending thus slows down dramatically and *thus* the buffer fills. ... TCP will packetise the data in buffer as suits the ... You have no direct control over what packets TCP will send ...
    (microsoft.public.win32.programmer.networks)
  • Re: UDP vs TCP
    ... TCP for instance will break up a large packet into smaller ... into the packets and then the receiving app would have to read ... Network Layer -> ethernet ... DOMAIN over port 53 ...
    (microsoft.public.vb.enterprise)
  • Re: serious networking (em) performance (ggate and NFS) problem
    ... > Desktop adapter MT) connected directly without a switch/hub ... network hardware at a lower level. ... If the problem is your connection is dropping a few packets, ... NFS over TCP does better than UDP does in lossy network conditions. ...
    (freebsd-current)
  • Re: serious networking (em) performance (ggate and NFS) problem
    ... > Desktop adapter MT) connected directly without a switch/hub ... network hardware at a lower level. ... If the problem is your connection is dropping a few packets, ... NFS over TCP does better than UDP does in lossy network conditions. ...
    (freebsd-stable)