Re: [NFS] nfs performance: read only/gigE/nolock/1Tb per day

[Bogdan Costescu <bogdan.costescu@iwr.uni-heidelberg.de>]

[ cc-ed to netdev; the discussion was about receiving bursts of ICMP Time 
Exceeded messages after some large NFS datagrams could not be reassembled; 
sometimes down/up the interface on the receiver/reassembly side cures it ]

On 23 Apr 2002, Trond Myklebust wrote:

>      > How big are the datagrams compared with the MTU ? With 32K
>      > datagrams over Ethernet, you're talking about roughly a full Rx
>      > ring worth of packets (32 is common for the Rx ring size)...
> 
> It has been a while ago (I've since mothballed the machine) but I saw
> it on a Pentium 90 with only 8k write sizes. 4k was fine, 8k gave
> avalanches.

IMHO you can't comletely eliminate hardware related problems: apart from 
having a slow CPU, some early PCI implementations were buggy (although you 
don't say if it's PCI or ISA and what's the link speed).

>      > Does the other side sees these messages ?
> 
> IIRC, yes, and the server was resending the datagrams. From the code,
> it looks as if there is no attempt to stop loopback situations
> occurring when this goes on:
> i.e. resending an ICMP when the server resends a datagram which times
> out again appears to be possible. This might be what was happening...

That's why I cc-ed netdev. My knowledge above the driver level is close to 
non-existant...

-- 
Bogdan Costescu

IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.Costescu@IWR.Uni-Heidelberg.De