Re^10: UDP connection

[BrowserUk]

You are correct. I must have misread the numbers.

It happens. Thank you for acknowledging it.

However, there is no problem. Here are a sample server and client that do bidirection communication over 4 socket pairs.

Hm. There are still problems you are not dealing with. I'm not talk ing about error handling or niceties here.

1. Your server is not acknowledging (responding to) the clients heartbeats per the OPs description:

   And once it receives a "heartbeat" (a byte containing "01") it sends a response to the IP and port the hb came from (byte containing "25").

   It may sound a minor omission, but (I believe) it does complicate the design of the client (and server) somewhat.

2. Your server is using the receipt of the heartbeat as a trigger for the start of server to client data flow.

   In the OPs description:

   • the hexdumps are only sent after the initial heartbeat.

     (Obviously, how else would the server know there was a client to send to:)

   • But they are not triggered by the heartbeat!

     He goes on to say:

     If there are test results ready on the server it sends them to ....

     Which suggests, though it doesn't actually state -- and he never came back to answer the question -- that the sending of the hexdumps is initiated by the server, when they become available; provided that is within 10 seconds of a heartbeat.

The significance of those (perhaps apparently small) differences, is that your client treats all inbound communications as data. The OP cannot do this as he has also to distinguish between the heartbeat response and the actual data.

The single byte packet size of the response should make that easy... except when you take the uncertain timing of the availability of the hexdump data into consideration.

Here is the black hole in the timing scenario I was trying resolve with the OP, which your client does not -- and (I believe) as written, cannot -- resolve:

time ascending relative
  ...
  0.000
   At this point, a client has (just) sent a HB, and the server has ac
+knowledged it.
   The server had no data to send
  ...
  9.99999... seconds expire

 The server discovers that it has hexdump to send and starts transmitt
+ing ...
 10.000   
 The client sends it next heartbeat and awaits the response.

The hexdump doesn't have to be large, a single packet coming available (at exactly the wrong time) will trigger the problem. The client is expecting a single byte response. The server hasn't seen the heartbeat, so it isn't going to send it until (at least) after it finishes sending the current packet. If, after sending the heartbeat, the client went into a read state for the single byte response, it will get the first byte of the data packet, and the rest will be discarded. This is the exact scenario that the OP described in his first post.

Using blocking reads -- as would normally be used in conjunction with threads as indicated by the OP -- the client cannot be in a read state in anticipation of the arrival of a data packet that could come at any time; and also send a regular heartbeat, because you cannot send() via socket, if that socket is currently in a recv() state. (At the same end!)

Using select obviates that, by avoiding entering the read state until it knows (by polling) that data is ready to be read. But that alone does not completely close the window for failure.

If the 10 second timeout at the server occurs exactly whilst the client is in the process of receiving the latest packet -- and is therefore unable to transmit its next heartbeat -- the server will (may) conclude that the client has 'gone away' and discard (fail to send) any subsequent data until the client re-heartbeats. And that leads to data loss, which the OP explicitly denounced.

Whilst the window for such a failure is small; such are the nature of communications protocol failures.

The usual solution(s) to this problem include:

1. Have the client send the heartbeat at twice the frequency of the servers timeout.
2. Have the server declare a timeout of half what it will actually accept.
3. Have the server reset the timeout whenever it sends -- be it heartbeat ack; or data packet -- rather than when it receives.

And it was these details I was trying to establish with the OP before he got scared away by our ...um... discussion.

[JavaFan]

Well, only the OP can answer all those questions.

I did ignore the single byte response on purpose; my main goal was to show bidirectional communication was possible, and there's no two processes trying to get a hold on the same socket.

As I understood the situation, everything the server sends is a hexdump, except the very first byte. Note that my client wasn't writing anything to a file; that, and leaving off the first byte I left as an exercise to the user. Not sending anything (or just only the single byte) after receiving the first heartbeat doesn't seem to complicate anything, IMO. Now, I am assuming (and I may be incorrect in that), an acknowledge is send only after the first heartbeat (in which it's safe, and easy, to just ignore the first response). I agree it becomes more complicated if the client cannot distinguish between heartbeat responses and hexdumps. But it may very well be that the nature of the hexdumps send is such that there will not be a hexdump consisting of a single byte with value 25.

If the 10 second timeout at the server occurs exactly whilst the client is in the process of receiving the latest packet -- and is therefore unable to transmit its next heartbeat -- the server will (may) conclude that the client has 'gone away' and discard (fail to send) any subsequent data until the client re-heartbeats. And that leads to data loss, which the OP explicitly denounced.

Yes, that's why my client sends a heartbeat every 9 seconds. ;-) (This is similar to solution 1) The select loop times out every second, after which at most 1 packet is read before checking the time, and a heartbeat is send if we're 9 seconds or more after sending the previous heartbeat (and this is easily tweakable). In theory, the client could be so busy that the process is swapped out for a full second. Or even more than 10 seconds, not having any opportunity at all to send a heartbeat.

Have the server reset the timeout whenever it sends -- be it heartbeat ack; or data packet -- rather than when it receives.

I'm not sure that would work. Say, the server has more data to send every second. If it would reset the timeout each second, it would never notice the client going away as the timeout keeps being pushed further back.

[BrowserUk]

my main goal was to show bidirectional communication was possible,

Bidirectional communication was never in question (by me).

My questions/concerns related to concurrent bidirectional comms (at the same endpoint), which AFAIK is impossible on any platform.

Of course, this never arises with the select loop mechanism, as there is only one thread of execution and it is either waiting; or processing a can_read state; or processing a can_write state; or an error state.

But with either a multi-threaded, or multi-process solution on multi-core hardware, it can be attempted -- as exampled by the idea that the client run one thread/process for doing the heartbeat and another for receiving the data -- and needs to be handled.

It can be handled (with threading at least), but unless the OP comes back and fills in a few of the blanks, there's probably little point in further pursuing the possibilities.

[falstad]

I'm sorry for ignoring the node for so long. I took some days off from work for an exam -.- I've read the whole discussion and I'd like to thank you for your replies.

As above mentioned the UDP packet size is never bigger than 65535 bytes, and the read won't last 10 secs either.

While sending out the hexdump the server will notice the heartbeat and will send a response after it sent the data. The server will always respond to the heartbeat. Also we know for sure, that there is no hexdump which contains only one byte. They have several types and the smallest is around 20 bytes.

The server can do with some delay in receiving the heartbeat. For testing it's fine to set up the timer to send heartbeat in every 9.9 secs.

"The sending of the hexdumps is initiated by the server, when they become available; provided that is within 10 seconds of a heartbeat." It's possible that we start the server and there is no hexdump ready. So it is possible that we have to wait for several minutes for the first hexdump to come. The client continues to send the heartbeats regardless of it receives no data.

I hope I've answered everything that wasn't clear.