Re: [collectd] zeormq architecture inquiry

2011-08-29 Thread Allan Feid
So I talked to some people in the #zeromq channel on freenode, and have been
reading through the zeromq guide. Seems like at scale, with tons of nodes,
you would need pretty fast disk I/O to handle all the writes to RRD. Using
ZeroMQ, you could dish out to different collectd worker machines to handle
the data of certain kinds, example CPU on one worker and memory on another,
which would allow you to distribute your disk i/o amongst smaller and
cheaper nodes, possibly VMs. That's a bit complex for right now, but an idea
worth thinking about in the future. This would involve a zeromq process that
routed requests to various worker machines.

I think for my initial attempt at this, I will have one collectd instance
set to subscribe, and each node set to publish to my subscriber's endpoint.
This should allow nodes to come/go as they please, hopefully this goes well,
as the flexibility in zeromq makes a great way to distribute work.

Thanks,
Allan

On Mon, Aug 29, 2011 at 11:58 AM, Florian Forster o...@collectd.org wrote:

 Hi Allan,

 On Sun, Aug 28, 2011 at 08:11:14PM -0400, Allan Feid wrote:
  I was wondering if anyone has any experience doing this, or can
  provide some guidance on what zeromq architecture would scale best.

 I was wondering the same thing. For a normal (intra-)DC setup, I think
 it the best option would be a central subscriber that binds to a local
 address and many publishers which connect to that address. For an
 inter-DC setup, I'd probably go with aggregators on a DC level and
 repeat the same pattern, i.e. publish to a remote address and, on the
 global level, subscribe to a local address.

 Disclaimer: I hardly have any ZeroMQ experience either, so this is
 basically how I think stuff should work, but I might be totally wrong
 about it …

 Best regards,
 —octo
 --
 Florian octo Forster
 Hacker in training
 GnuPG: 0x0C705A15
 http://octo.it/

 -BEGIN PGP SIGNATURE-
 Version: GnuPG v1.4.10 (GNU/Linux)

 iQIcBAEBAgAGBQJOW7cQAAoJEMPSHpbi2Mmg9PAP/AtwcBZdyS7qDPZC57dodFnV
 Kaj3Z5k5TJqzVE7+n5JeVYrMkCUelHKI7LTovQvOZ8m/ViYz3tOugfAoRr4pU3M5
 GYflMXWAW5yES0+srZsTxChu22n717zwnL3DjqM4byZxL2b0S4h5I7v5pyWqqYSM
 LmZOOXwy/H8mWXcl1qsMgFBeV0s8jQnITq/yRjx+sQcMDk3Pj094AmLPSNh3tLrC
 mipOfWyXWBcsFtr7x9K+sqY6PFxmpZEJYU5JEJG+pptRdNmOvPcfGxjlkB4b1MQ3
 SlpCLHsWY28/e3fwQ0RVK0dc6+PqzqRLocCRxphbdqRYlZ3EJ9m4Zq+rMtrHra+r
 i4SpmptQQgRvp7vftv1tE3vYh7jCa7a3li7KhOA7axS1IHQh4zasN/c5/CGnXDLE
 RvaW4g71xOOoWe63SRABEOUiY+AHRS5+ZiY/+bprscofbd+HmE7CiFK/ZGEMnuG+
 eRX+L4sm7dhYckGe7AnokW5fVKOzJd9kJZ36nhEZOUFpLiGIDjoieRK9YZtVeVBF
 VIZV75WWElT+zFOUcG8tFulR4QrYlsRzoAs6AnvAMRZ4YjfaaIE5IS7Q+GbeQCcQ
 YuzsXm6E6h/ifx+gBtChpUpAEc+qanS1q1Y8RHCwP7RBEr8mxAYgS4v/YKBV62M9
 7Pckd9P+aN6HuHRL/PUN
 =6w1f
 -END PGP SIGNATURE-


___
collectd mailing list
collectd@verplant.org
http://mailman.verplant.org/listinfo/collectd


Re: [collectd] zeormq architecture inquiry

2011-08-29 Thread Florian Forster
Hi Allan,

if you're after distributing the load, you might want to take a look at
the AMQP plugin, too. It has these topic branches, where subscribers
can subscribe to matching messages only. This allows you to distribute
the load in a consistent manner and is re-configurable during runtime.

It's possible to only subscribe to certain messages in ZeroMQ as well,
but I *think* they match the message content for this. Since the plugin
uses the binary protocol, that might not work for us. I might be wrong,
though, and there is some label that you can attach to messages, which
would make this kind of filtering possible with ZeroMQ, too.

Regards,
—octo
-- 
Florian octo Forster
Hacker in training
GnuPG: 0x0C705A15
http://octo.it/


signature.asc
Description: Digital signature
___
collectd mailing list
collectd@verplant.org
http://mailman.verplant.org/listinfo/collectd


Re: [collectd] zeormq architecture inquiry

2011-08-29 Thread Francois-Xavier Bourlet
Hi guys,

Just a little note: The current implementation of ZMQ filter on the
subscriber, not on the publisher...
Seem that ZMQ v3 is addressing the problem.

my 2cents

On Mon, Aug 29, 2011 at 9:22 AM, Florian Forster o...@collectd.org wrote:

 Hi Allan,

 if you're after distributing the load, you might want to take a look at
 the AMQP plugin, too. It has these topic branches, where subscribers
 can subscribe to matching messages only. This allows you to distribute
 the load in a consistent manner and is re-configurable during runtime.

 It's possible to only subscribe to certain messages in ZeroMQ as well,
 but I *think* they match the message content for this. Since the plugin
 uses the binary protocol, that might not work for us. I might be wrong,
 though, and there is some label that you can attach to messages, which
 would make this kind of filtering possible with ZeroMQ, too.

 Regards,
 —octo
 --
 Florian octo Forster
 Hacker in training
 GnuPG: 0x0C705A15
 http://octo.it/

 -BEGIN PGP SIGNATURE-
 Version: GnuPG v1.4.10 (GNU/Linux)

 iQIcBAEBAgAGBQJOW7y2AAoJEMPSHpbi2Mmg0ysP/ijjRkTLWdvFCxs+sBBXtvU3
 liEIgNf90s3Jgzr33bgx+7m/BQ0JgDT7VrNvH3ek8/OdePGJ15Lnnqr6T5mDz2gD
 Tj0uysiFCyQCIZPiw5v23x+iNLhwkeS/RC9HFBTPTG3I+s4jG46EvVvBfjlosEYA
 g+LslGIZ7yG02PN/WKxehYUEK3nnWlz14uHyFwPoQfUGWObpQiiKzerM5R0uL/qn
 rwyK+IFlmAgzU0m1cbaGq3KCyqOGDHzreW6rC1fql8IFC1ca2ugFA3F6ttYUzRvV
 yP1RYwydWwA6WYjYyyExWsSTQTWhIyx0JWnxZslodhdA2pEHU+OGsciuPoBnCW5c
 35GAozEiAuRFe2QKFPnhDkLyVf8Wh/agyWfCSR8Q1i2UVVrhWUFRIaOCzC/+e8od
 UamMKv3KkuC931M6YDbLNW4W/E71rQB9sAxCOItVmOpISaW5nCMiw59dknZwN2I4
 WsdmTYzMdQZhmjB28/i1fo0vZVuWHH/dK0zAoaWwutanthlVFKrSxcnmfkjzQNj1
 WlHqXDvW6h7crksOeWQamrUqplHgP0sKSWsTI2z4zytCGL3+KpdeXe/0FBqPou7J
 qHqFM2nTKdRNpL8fgcRm9nTz0VyB0vmpmrIQ8WgOjpwfNz9APUrW5s1RECovb+ow
 4wLyW448ylCs6WC7HXwE
 =HGPq
 -END PGP SIGNATURE-

 ___
 collectd mailing list
 collectd@verplant.org
 http://mailman.verplant.org/listinfo/collectd




-- 
François-Xavier Bourlet
___
collectd mailing list
collectd@verplant.org
http://mailman.verplant.org/listinfo/collectd


Re: [collectd] zeormq architecture inquiry

2011-08-29 Thread Allan Feid
Thanks Francois, the way the code is right now, the subscriber is the node
doing the bind, and the publisher connects to the subscriber. This would
mean that your central collectd process would be able to do the filtering
since it is the subscriber.

As far as AMQP goes, we have noticed scaling problems with the broker model,
where this box becomes complex to scale out and seems to be the bottle neck.
That is why ZeroMQ is appealing for this kind of situation.

On Mon, Aug 29, 2011 at 12:42 PM, Francois-Xavier Bourlet
f...@dotcloud.comwrote:

 Hi guys,

 Just a little note: The current implementation of ZMQ filter on the
 subscriber, not on the publisher...
 Seem that ZMQ v3 is addressing the problem.

 my 2cents

 On Mon, Aug 29, 2011 at 9:22 AM, Florian Forster o...@collectd.orgwrote:

 Hi Allan,

 if you're after distributing the load, you might want to take a look at
 the AMQP plugin, too. It has these topic branches, where subscribers
 can subscribe to matching messages only. This allows you to distribute
 the load in a consistent manner and is re-configurable during runtime.

 It's possible to only subscribe to certain messages in ZeroMQ as well,
 but I *think* they match the message content for this. Since the plugin
 uses the binary protocol, that might not work for us. I might be wrong,
 though, and there is some label that you can attach to messages, which
 would make this kind of filtering possible with ZeroMQ, too.

 Regards,
 —octo
 --
 Florian octo Forster
 Hacker in training
 GnuPG: 0x0C705A15
 http://octo.it/

 -BEGIN PGP SIGNATURE-
 Version: GnuPG v1.4.10 (GNU/Linux)

 iQIcBAEBAgAGBQJOW7y2AAoJEMPSHpbi2Mmg0ysP/ijjRkTLWdvFCxs+sBBXtvU3
 liEIgNf90s3Jgzr33bgx+7m/BQ0JgDT7VrNvH3ek8/OdePGJ15Lnnqr6T5mDz2gD
 Tj0uysiFCyQCIZPiw5v23x+iNLhwkeS/RC9HFBTPTG3I+s4jG46EvVvBfjlosEYA
 g+LslGIZ7yG02PN/WKxehYUEK3nnWlz14uHyFwPoQfUGWObpQiiKzerM5R0uL/qn
 rwyK+IFlmAgzU0m1cbaGq3KCyqOGDHzreW6rC1fql8IFC1ca2ugFA3F6ttYUzRvV
 yP1RYwydWwA6WYjYyyExWsSTQTWhIyx0JWnxZslodhdA2pEHU+OGsciuPoBnCW5c
 35GAozEiAuRFe2QKFPnhDkLyVf8Wh/agyWfCSR8Q1i2UVVrhWUFRIaOCzC/+e8od
 UamMKv3KkuC931M6YDbLNW4W/E71rQB9sAxCOItVmOpISaW5nCMiw59dknZwN2I4
 WsdmTYzMdQZhmjB28/i1fo0vZVuWHH/dK0zAoaWwutanthlVFKrSxcnmfkjzQNj1
 WlHqXDvW6h7crksOeWQamrUqplHgP0sKSWsTI2z4zytCGL3+KpdeXe/0FBqPou7J
 qHqFM2nTKdRNpL8fgcRm9nTz0VyB0vmpmrIQ8WgOjpwfNz9APUrW5s1RECovb+ow
 4wLyW448ylCs6WC7HXwE
 =HGPq
 -END PGP SIGNATURE-

 ___
 collectd mailing list
 collectd@verplant.org
 http://mailman.verplant.org/listinfo/collectd




 --
 François-Xavier Bourlet

___
collectd mailing list
collectd@verplant.org
http://mailman.verplant.org/listinfo/collectd


Re: [collectd] zeormq architecture inquiry

2011-08-29 Thread Shanker Balan
Hello,

Allan Feid wrote,
 Hello,
 
 I've been looking into using collectd combined with zeromq for transport.
 This seems like a fairly good way to scale out monitoring infrastructure as
 each node being monitored can send out requests over zeromq to one or more
 locations with little overhead. I was wondering if anyone has any experience
 doing this, or can provide some guidance on what zeromq architecture would
 scale best.
 
 With my limited zeromq knowledge, I'm thinking that each node being
 monitored would publish to one or more collector nodes, which could then
 reroute the information down to a single point where graphing and alerting
 thresholds could be handled. The collector nodes would likely be in each
 datacenter and running collectd to write to RRD files. These instances of
 collectd, could also be publishing or pushing data down to a
 single aggregation system. I guess my real question here, is what zeromq
 pattern would make most sense for a distributed collectd monitoring system.
 Any insight or suggestions are greatly appreciated.

My team initially started out with collectd but we ran into a requirement
where we needed to have the app itself emit events in real time. It was not
clear if the collectd libraries could be used within Java, Perl/Ruby etc
applications.

After a bit of grepping around, we settled on MonDemand with LWES
(UDP+Multicast) as the transport. ZeroMQ was an option, but we liked what we
saw in MonDemand.

If the need arises in the future, we plan to replace LWES with a MQ system.

lwes.sf.net
mondemand.sf.net

Hth.

-- 
http://shankerbalan.net/

___
collectd mailing list
collectd@verplant.org
http://mailman.verplant.org/listinfo/collectd