dnsdist: is my usage of newServer() correct or there is a better way?

[ousatov-ua]

Hi!

I have a pdns-recursor in recursion only mode, it runs with next config (I have cpu 16 threads):

tcp-threads=15
threads=15

As facade for it I have running dnsdist with config:

for i=1,15 do
    addDOHLocal()
    addDOH3Local()
    addTLSLocal()
    addDOQLocal()
    
    # the questions about this downstream:
    newServer({address="127.0.0.1:5353"})
end

Is this newServer()'s is correct usage or I should play with sockets config for newServer() directive?

Sorry if the question is silly, I'm not a clever man.

And do you plan to support downstream newServer()'s QUIC?

Thank you in advance!

P.S. Many thanks for contributors for such a great services created!

[rgacogne]

It depends on how many CPU cores you want to dedicate to DNSdist, mostly. Every single newServer directive will start a new thread to deal with the responses from the Recursor, so starting 15 of them can use up to 15 CPU cores. For each directive there is only one socket between dnsdist and the recursor (so one <destination address, destination port, source address, source port> tuple).
If you instead use one newServer with sockets=16, for example, it will start only one thread to deal with incoming responses, but it will use 16 sockets between dnsdist and the Recursor. Increasing the number of sockets might make sense if you are enabling reuseport in the recursor configuration, because Linux doesn't load-balance incoming packets very well in that case. We have written more words about that in https://dnsdist.org/advanced/tuning.html so I suggest you start by reading this page, and ask any question you might still have after that.

And do you plan to support downstream newServer()'s QUIC?

If there is actual interest, yes. We have not seen a lot of interest so far, and if I have to guess I suspect people will be more interested in support for outgoing DoH3 than for outgoing DoQ, but I might very well be wrong.

[ousatov-ua]

Many thanks for your answer!
Yes, I see - one thread per one declaration. :)
I will read documentation.

Maybe it is not related to dnsdist (but to Linux itself), but I want to ask you one more question if you don't mind:

In general, I want to utilize all my CPU for dnsdist

How much good dnsdist works in HT, so let's say I have 8 physical core and 16 with HT.
Will it give so much benefit (as 2x) to use exactly 16 instances vs 8 instances of dnsdist?

Thank you!

[ousatov-ua]

No needs to answer, I've read documentation.