Your number of resources and time of your catalog runs

[Bostjan Skufca]

Hi,

I was just wondering what is someone else's number of manager resources and catalog run time.

Mine is 1,700 resources and average catalog run of 70-100 seconds (40 seconds for catalog retrieval, 30-60 for catalog run), depending on the network distance between puppetmaster and managed node.

Then there are few more seconds for provider cache stuff, and report uploading takes it's tool too, together adding some 20 seconds.

Care to share your puppet agent performance details?

b.

[Jakov Sosic]

We're not as near in the number of resources per node yet, so no point
in comparing :) Let's just say my client run is less then 10 seconds on
average :)


But as a side note, I'm wondering at what number of nodes one must
replace puppetmaster with Apache+Passanger? Just a rough estimate, let's
say master runs on 2 socket quadcore.

Is it 50 agents? 100 agents? 200 agents?

[llowder]

I replaced it at around 35 nodes I think. I've currently got 42. Before making the change to Apache + Passenger, we were seeing timeouts with webrick.

And in all honestly, I would make that change if you were using puppet in any sort of production environment. The only time I use webrick is for throwaway VMs where I have 1-3 nodes (including the master itself) running off a master so I can test something outside of my main master.

 

[Ramin K]

On 10/22/2012 2:11 PM, Jakov Sosic wrote:
>
> But as a side note, I'm wondering at what number of nodes one must
> replace puppetmaster with Apache+Passanger? Just a rough estimate, let's
> say master runs on 2 socket quadcore.
>
> Is it 50 agents? 100 agents? 200 agents?
>

Available CPU has no bearing on your question. Webrick is single
threaded so the limiting factor is concurrency > 1. If your client
requests aren't overlapping then webrick will work fine. However if
you're using Puppet to support production servers I'd use
Passenger/Apache regardless of whether you're likely to hit the limits
of webrick.

Ramin

[Bostjan Skufca]

Hi Jakov,

On 22 October 2012 23:11, Jakov Sosic wrote:

We're not as near in the number of resources per node yet, so no point

in comparing :) Let's just say my client run is less then 10 seconds on
average :)

Those were the nice old days, when puppet catalog was still small and it run fast :)

But as a side note, I'm wondering at what number of nodes one must
replace puppetmaster with Apache+Passanger? Just a rough estimate, let's
say master runs on 2 socket quadcore.

Is it 50 agents? 100 agents? 200 agents?

I'm thinking about it now, but at the distributed deployment, where each puppetmaster serves around 20-30 nodes it is really not worth a hassle by itself, but I will do it soon anyway, curiosity reasons. Also, nodes connect to master only once every hour, seems often enough.

b.

[Jakov Sosic]

Thank you very much for your information, (both llowder and you).

I will migrate to Passanger ASAP. :)