[erlang-questions] UDP concurrent server

[Bengt Johansson E]

Hi Bogdan!

I’m sure you understand the difference between TCP and UDP, I just wanted to highlight it for the discussion. ☺

Given that I don’t know your application, it is difficult to give specific advice. But let’s assume that your application involves sending messages using an application specific protocol to your server. Let’s also assume that each client needs to communicate with the server more than once, so we essentially have a connection between the clients and the server.

In that case you will have a dispatcher process like you mention that sends packets along to the processors. This could be round-robin, random or based on some identifier in the packet. In fact, the concurrent TCP server is implemented using that pattern as well, but it hands off the classification of connections to the TCP/IP stack by using different ports for each connection.

I agree that it would be kind of nifty for some applications to allow the UDP stack to spread incoming packets to more than one receiver, but that would be limited to random or round-robin distribution or some similar simple method since it doesn’t understand the contents of the packet.

Unfortunately, the dispatcher “pattern” puts a cap on the maximum possible concurrency. Amdahl’s law states that the maximum possible concurrency in a parallel system is limited by the sequential parts of the system since they will be saturated and then the parallel parts will just idle. So if the time spent in the sequential part (the dispatcher) is 10% per packet then you will get a maximum 10-fold increase in speed if you go parallel. If you spend 1% there you will get a 100-fold increase and so on.

Assuming that you application does a lot of computing for each packet (or waits for disk etc.) you are likely not going to run into that problem, but if you are implementing something that does very little to the packets (like a router or firewall) this might be a problem for you.

Note that I reason in terms of processing time, I don’t see why you can’t have millions of open connections if they do not communicate too much. Erlang is an excellent language for that kind of applications.

As for over-load situations, the UDP stack will handle that for you by dropping packets when it’s buffers run full. If that is a problem your application protocol will have to handle packet retransmits by itself.

Good luck with your project!

BR, BengtJ


From: Bogdan Andu [mailto:bog495@REDACTED]
Sent: den 9 december 2015 12:00
To: Bengt Johansson E
Cc: Erlang
Subject: Re: [erlang-questions] UDP concurrent server

Hi Bengt,
Yes I am aware that udp sockets are working different than tcp sockets,
but I was wondering if there is an idiom for udp as it is for tcp to make things concurrent.
Having in mind the limitations of udp it seems that the best option
is to implement the version with a controller process that receives udp packets
and in spawning processes to actually handle that packet based some info, and immediately
fetch the next message from message box.
My only concern here is the bottleneck. The message box can easily be overloaded
and the response of the server exponentially delayed.
Is there any best practices to handle such situations avoiding message box overloading
while handling say 1 million of concurrent (udp) connections ?
Thank you,
/Bogdan



On Wed, Dec 9, 2015 at 12:35 PM, Bengt Johansson E <bengt.e.johansson@REDACTED<mailto:bengt.e.johansson@REDACTED>> wrote:
Hi Bogdan!

Your questions make me a bit confused. I wonder *why* you want two processes waiting for packets from the same socket and *what* you expect to happen?

Generally one usually only speaks about concurrent servers regarding TCP where you bind to a socket waiting for incoming connections and once a connection is established, you spawn a process (or thread depending on language) and process the data coming over the connection concurrently.
Note that the main loop of the server that waits for responses is always sequential. The new process handling the connection gets a new socket with a free port used only for that particular connection.

But! UDP lacks support for connections mainly since it is a message based protocol and hence is devoid of any connection abstraction ☺

Still the question remains what is a concurrent UDP server? I guess what you want to achieve is some kind of distribution of incoming packets to several processes to handle them. In that case you should either go for the solution of TCP to set up new socket for each communication or write a simple process that classifies the incoming packets and distributes them to the correct process based on some information in the packet, fi. Some identifier you have chosen to identify the connection. Basically  you have to implement an application specific load balancer – or rather load distributor – hoping that the time taken to actually process the packets greatly overshadows the time it takes to distribute them. ☺

Anyway, there is no way the underlying UDP stack and/or erts can make the decision for you. To which process to send the packet that is.

Hope that helps!
BR, BengtJ






From: erlang-questions-bounces@REDACTED<mailto:erlang-questions-bounces@REDACTED> [mailto:erlang-questions-bounces@REDACTED<mailto:erlang-questions-bounces@REDACTED>] On Behalf Of Bogdan Andu
Sent: den 9 december 2015 10:15
To: Erlang
Subject: [erlang-questions] UDP concurrent server

following the thread https://groups.google.com/forum/?hl=en#!topic/erlang-programming/6Q3cLtJdwIU
as it seems that POSt to topic does not work

After more tests the basic questions that remains ..

Is there a way to have more than one process be blocked
in gen_udp:recv/2 call as this seems to not be possible,
probably because the way udp sockets work.

Sequentially works as expected, but when when I try to spawn another process
that makes and attempt to execute gen_udp:recv/2 while the first process already does
gen_udp:recv/2 , the second process gives elready error. This means that 2 process
cannot concurrently do gen_udp:recv/2 .

In scenario with socket {active, once} or {active, true} there is only one process that can
receive the message from socket (the one that does gen_udp:open/2 ),
and for multi-threaded applications this quickly can become a bottleneck.
In this case, however, elready error disappears of course.
.
I tried both variants and both have disavantages.

Is there an idiom for designing a udp concurrent server in Erlang?
So far, it seems impossible.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://erlang.org/pipermail/erlang-questions/attachments/20151209/ce28b5dc/attachment.htm>