Re: Kernel 2.4.6 & 2.4.7 networking performance: seeing serious delays inTCP layer depending upon packet length

[Ron Flory <ron.flory@adtran.com>]

Shirley Ma wrote:
> 
> Hi, Ron,
> 
>      I am interested the problem you posted. I tried to reproduce
> this problem on my machine and failed. Would you please point your 
> programs to me, so I can reproduce this problem and do more 
> investigation? If not, please let me know whether it is reproducible,
> and please collect both client/server ethereal log.

 OK, will follow-on in a private email (to minimize lkml traffic)
 
 I can easily reproduce the problem by performing a socket 'write' as
two separate operations:

   write(sock, buf1..)
   write(sock, buf2..)

 If on the other hand I combine the buffers then issue a single
'write':

   write(sock, both_bufs...)

 the problem magically disappears (because the inter-block handshaking 
requirements change, which is where I think the problem actually lies).

 Since the problem is also present using the loopback device, both 
client/server sides would be present in a Ethereal long of LO.

 I would imagine anybody running a large Linux ftp/http server would be
interested in this...

ron