[prev in list] [next in list] [prev in thread] [next in thread] 

List:       quagga-dev
Subject:    [quagga-dev 9495] Re: [RFC] AgentX support for Quagga
From:       Renato Westphal <renatowestphal () gmail ! com>
Date:       2012-06-26 20:36:33
Message-ID: CAChaegmEOESOafmm40DuX==8xbbpugbMQbcTbxS8bwheWJX__w () mail ! gmail ! com
[Download RAW message or body]

2012/6/26 David Lamparter <equinox@opensourcerouting.org>:
> On Tue, Jun 26, 2012 at 02:47:52PM -0300, Renato Westphal wrote:
>> 2012/6/26 David Lamparter <equinox@opensourcerouting.org>:
>> > On Tue, Jun 26, 2012 at 08:06:12AM -0700, Stephen Hemminger wrote:
>> >> On Tue, 26 Jun 2012 11:01:18 -0300
>> >> Renato Westphal <renatowestphal@gmail.com> wrote:
>> >>
>> >> > 2012/6/25 Vincent Bernat <bernat@luffy.cx>:
>> >> > > The upgrade is already done. The specific code to each daemon is
>> >> > > compatible (and already was) with both SMUX and AgentX (except the trap
>> >> > > part but I have already upgraded it). This is also why SMUX/AgentX is a
>> >> > > compile time option.
>> >> >
>> >> > Thanks for the clarification. Everything makes sense for me now.
>> >> >
>> >> > Just one note. I have to run Quagga (any daemon) with root privileges
>> >> > in order to use agentx. Otherwise I get an "snmp[warning]: Warning:
>> >> > Failed to connect to the agentx master agent ([NIL]):".
>> >> >
>> >> > > AgentX won't improve the situation from the performance point of view.
>> >> > >
>> >> > > For bgpd, I suspect the problem to be fairly simple: snmpwalk is a serie
>> >> > > of GETNEXT requests. For each GETNEXT request, we need to locate the
>> >> > > appropriate element to return and for this, we need to walk each route
>> >> > > table from the beginning.
>> >> > >
>> >> > > An easy fix would be to remember where we were the last time we got a
>> >> > > GETNEXT.
>> >> >
>> >> > I don't think that's the problem. I did some profiling with
>> >> > callgrind[1] and found out that the libnetsnmp is inherently slow. I
>> >> > did a snmpwalk over a table with 10k routes and only a small
>> >> > percentage of the execution time (~ 1.3%) was to gather information
>> >> > from the BGP table (in the bgp4PathAttrTable function). I might be
>> >> > wrong, but I think there's nothing we can do about this problem.
>> >> >
>> >> > [1] http://www.inf.ufrgs.br/~rwestphal/files/callgrind.png
>> >> >
>> >> > > I will try to work on this. Is there some way to get a huge BGP
>> >> > > table (with exabgp maybe?)?
>> >> >
>> >> > I'm using the bgp_simple.pl script. I followed this tutorial to get
>> >> > started with it:
>> >> > http://evilrouters.net/2009/08/21/getting-bgp-routes-into-dynamips-with-video/
>> >> >
>> >> > > You are right, I didn't retest SMUX when I added trap support. I have
>> >> > > added your patch and rebased my stack of patches (to make all patches
>> >> > > compile, since David did not merge them yet). I have therefore "broken"
>> >> > > your tree. You can get it right with something like (assuming the remote
>> >> > > name for my repository is "rvincent")
>> >> > >
>> >> > > git fetch rvincent
>> >> > > git rebase --onto rvincent/feature/agentx vincent~1 vincent
>> >> > >
>> >> > > (or you can just start a new branch and cherry-pick your last commit)
>> >> >
>> >> > Perfect. I hope your patches will be merged in a few days.
>> >> >
>> >> > > Your patch seems fine. I don't have a testbed to actually test it but I
>> >> > > don't see any problem with it.
>> >> >
>> >> > Thanks for looking into it. I'll review this patch next week and then
>> >> > send it to quagga-dev.
>> >> >
>> >> > Regards,
>> >>
>> >> I did some work to speed up SNMP and it's handling of large tables.
>> >> A lot of the problem was N^2 insertion into tables, combined with cache
>> >> timeout being less than the load time.
>> >
>> > Do these changes conflict with Vincent's AgentX support?  Could you
>> > (re-)post them to the list?
>> >
>> > Thanks in advance,
>> >
>> > -David
>>
>> His changes are in the net-snmp, not quagga:
>> http://git.vyatta.com/git/?p=net-snmp.git;a=shortlog;h=refs/heads/oxnard
>
> Oh, ah.
>
>> I'm testing them now.
>
> Perfect, thanks!

Well, I don't know if I did something wrong but I didn't notice any
performance improvements after applying Stephen's patches on net-snmp.
I'll take a better look at it tomorrow..

-- 
Renato Westphal

_______________________________________________
Quagga-dev mailing list
Quagga-dev@lists.quagga.net
http://lists.quagga.net/mailman/listinfo/quagga-dev
[prev in list] [next in list] [prev in thread] [next in thread]