<div>The step 2 is the consequence of the naive notification of the worker from the facade. Every time a facade receives a request it notifies the worker there is something. Imagine 1000 requests arrive while the worker is working. When the worker performs its current job, it will take the next message, and pull all data from the facade. Then it will do 999 of "dummy" processing, because all the work was already done in the processing of the 1st message.</div>
<div>That is why, in my implementation, the worker flushes all incoming messages, then pulls all data from the facade.</div><div><br></div><div>Admittedly, this could be avoided by introducing a more elaborate communication between the worker and the facade. In my mechanism, the facade is not aware whether the worker is doing some processing or not, so it blindly sends it a message each time it receives a processing request.</div>
<br><div class="gmail_quote">On Sat, Feb 16, 2013 at 5:58 AM, Erik Søe Sørensen <span dir="ltr"><<a href="mailto:eriksoe@gmail.com" target="_blank">eriksoe@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<p dir="ltr">I must confess that (at least at this late hour) your Step 2 confuses me.<br>
Why not just 1+3+4, or indeed *just* step 2?</p>
<p dir="ltr">Regards, <br>
/Erik</p>
<div class="gmail_quote">Den 15/02/2013 16.47 skrev "sasa" <<a href="mailto:sasa555@gmail.com" target="_blank">sasa555@gmail.com</a>>:<br type="attribution"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div><div class="h5">
Hello,<div><br></div><div>A while ago I encountered the following situation:</div><div><br></div><div>I had the gen_server base process P which would receive messages, and handle them by sending some data over the network. The messages were coming faster than they were being sent. I established the reason for this was the "randomness" of my network conditions. I also established that sending more messages at once was almost as fast as sending one message, i.e. the network push time wasn't highly dependent on the message size.</div>
<div><br></div><div>To tackle this in a generic way, I devised an approach which has served me well in multiple places. I was repeatedly googling whether some similar solution exists, but I couldn't find it. Now, I'm not sure if I have reinvented a wheel, or the approach is not optimal, so I'm asking if you are aware of similar approaches, and are there any faults in this one?</div>
<div><br></div><div>The approach I took is following:</div><div><br></div><div><div>I split the server in two processes: the "facade" and the worker. The facade acceptes requests from clients, and stores them internally. While the worker is doing its work, new messages are stored in the facade. When the worker is available, it will take all accumulated messages from the facade and process them.</div>
</div><div><br></div><div>These are the steps:</div><div>1. The facade receives messages, stores data in its list, and notifies the worker (without sending actual data), that some work is ready.</div><div>2. Upon receiving the notification, the worker first flushes its message queue by doing repeated receive ... after 0 as long as there are messages in the queue.</div>
<div>3. Then the worker pulls all messages from the facade. This is a gen_server:call to the facade which will return all messages, and at the same time remove them from its state.</div><div>4. Finally, the worker processes all messages.</div>
<div><br></div><div><br></div><div>I found this approach useful because the delay on the worker adapts to the incoming message rate. </div><div>If the worker can handle messages at the incoming rate, everything works without delay. </div>
<div>If messages can't be handled at the incoming rate, the worker's delay will increase to accomodate the higher load. In other words, the worker will try to compensate the load by bulk processing messages. Obviously, this is useful only when process_time(N messages) < N * process_time(1 message).</div>
<div><br></div><div>Another benefit I found is that I can vary the implementation of the facade i.e. I can store messages using different algorithms. In the first implementation, I stored messages in a list. In one variation, I used hash which allowed me to eliminate duplicate messages. Another variant was truncation of the list, which allowed me to discard old messages if the queue was getting too large.</div>
<div><br></div><div>As I said, this has served me well in the production for more than a year, and I have finally found the time to make a generic library out of it. Before putting it public, I'd like to check if there are similar solutions, or alternative approaches?</div>
<div><br></div><div>Thanks, and best regards,</div><div>Sasa</div>
<br></div></div>_______________________________________________<br>
erlang-questions mailing list<br>
<a href="mailto:erlang-questions@erlang.org" target="_blank">erlang-questions@erlang.org</a><br>
<a href="http://erlang.org/mailman/listinfo/erlang-questions" target="_blank">http://erlang.org/mailman/listinfo/erlang-questions</a><br>
<br></blockquote></div>
</blockquote></div><br>