[lkcd-devel] RE: [lkcd-general] netdump throughput

[Deepak K. G. N. <dk...@no...>]

Hi Rajeev
> 1) Are there any published numbers or has anyone performed 
> similar tests
> and see the same kind of  numbers ? (just to make sure we arent doing
> anything
> wrong there)

You are right, the dumping mechanism is slow, i have tested it at my end. In
the early development stage there was a problem of incomplete dump and in
order to fix that problem i have provided a temporary patch which have slow
down the dumping mechanism. Also i have mentioned that even better solutions
are possible. See the snip of my mail on LKCD mailing list..

+++++++++++snip of my mail at LKCD mailing list ++++++++++++++++++++++++

Hello Everybody

I am sending the TEMPORARY solution (changes in Intel's solution) for having
the complete dump over the network. Using this code, I am able to take the
complete dump. It should be noted that this solution has PERFORMANCE ISSUES
and __even better solutions are possible__. I am currently __working__ on
them.

In brief, I have observed that in function netconsole_netdump (defined in
file netconsole.c) the delay was very small (while waiting for new request
from network dump server). I have increased the delay (from 100 micro second
to 1 millisecond).

It should be noted that, I have done very limited testing for this solution.

Precious guidance/suggestions/comments from experts are most welcome.

Patch Details:
==============

The patch contains following: -

1. netconsole_21-oct.tar.gz: This contains the netconsole module with
changes for increased delay.

2. netdump-21-oct: This is the initialization script (to be placed as
/etc/init.d/netdump)with some modifications. This script should be copied to
/etc/init.d/netdump.

Thanks 
Deepak Kumar Gupta
HCL Technologies Limited
Noida U.P.- INDIA

+++++++++++++++++++++++ end of snip+++++++++++++++++++++++++++++++++++++
> 2) The low bandwidth maybe due to stop-and-wait protocol with netdump
> server.
> I have noticed that after sending every page of data, we 
> perform a handshake
> with the server (send a hello, get an ACK and then go and 
> dump the next
> page).
> This would be one of the factors in bringing down the link 
> utilization and
> hence
> contributes to delay in dumping. Enhancing this to dump a lot 
> more data and
> support
> out-of-order could result in faster coredumps over the network.
> If yes, are there any plans to enhance it to support such a scheme ?

Here the aim is to have _complete_ dump _reliably_, thats why stop and wait
sort of mechnanism is used. However i am not clear which kind of
out-of-order support you are talking about ? Can you clarify me the
mechanism you are talking about ?

> 3) Is someone actively working on such an enhancement ?

Yes, i am working on it in my spare time .. not that much actively.. 
> Thanks
> Rajeev
> 

Thanks Rajeev
Deepak.