lkml.org 
[lkml]   [1999]   [Feb]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: Linux 2.2.2 TCP delays every 41st small packet by 10-20 ms
Our MPI-based applications involve communicating processes where message
delivery times have to be on the order of 100-200 MICROseconds, not 200
milliseconds. Therefore, Nagle is turned off via the TCP_NODELAY socket
option, and switched fast ethernet is used. Data aggregation suitable
for a fast typist is an impediment to MPI, where delays must be 1000
times shorter.

Sincerely,
Josip

Rogier Wolff wrote:
>
> David Miller wrote:
> >
> > 2) Even if your fix was to be the final one, we have to be careful,
> > as just blindly ACK'ing immediately for small PSH packets is
> > deadly and will kill Nagle immediately. Actually the effect is
> > that nobody will be able to successfully perform Nagle to us.
>
>[...]
>
> I didn't look up RFC 896, but Richard steven says:
>
> We saw in 19.3 that the round trip time on an ethernet for a single
> byte to be sent, acknowledged, and echoed averaged around 16 ms. To
> generate data faster than this we would have to be typing more than 60
> characters per second. This means that we rarely encounter this
> algorithm when sending data between two hosts on a LAN.
>
> (TCP/IP illustrated, (first or second edition) page 267. section 19.4)

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:50    [W:0.029 / U:0.684 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site