[Info-vax] counting-lock ?

[Jeremy Begg]

Hi Pierre,

A lot of discussion about how to implement a queuing mechanism using 
distributed locks ... but perhaps you don't need to do it that way at all.

It sounds like your DECnet objects use transparent task-to-task 
communication: DECnet creates a NET$SERVER process for each incoming 
connection, and the process runs your database query.  This is simple to set 
up and simple to program, but gives you very little control.

You might be better off by restructuring your application so that the server 
uses "non-transparent" communication.  In this mode, the server process 
starts up before any client requests have begun.  The server declares itself 
to DECnet as a network object, and then waits for incoming connections. 
Under this model your server can queue the work to be done and assign it to 
any number of worker processes, all under the control of the central server.

Have a look at the "DECnet for OpenVMS Networking Manual" which you can 
download from the HP OpenVMS web site
http://h71000.www7.hp.com/doc/73final/documentation/pdf/DECNET_OVMS_NET_MAN.PDF

Regards,

	Jeremy Begg

Pierre wrote:
> hi,
> 
> I have a DECnet object which can serve various known queries on a
> database. these queries are resource consuming and I don't want more
> than (say) n=5 of them of the same type to run at the same time. I
> want the other one to be queued up for further processing.
> 
> I first thought of using the lock manager. this would have work with
> an exclusive lock for n=1 but n>1...
> 
> I then thought of using counting-semaphores but the problem with
> semaphores is that if the server process unexpectedly die before
> signaling the its done with the resource, the slot it was using is
> never freed.
> 
> as unfortunately the DLM does not have "counting-locks" I had the idea
> of creating a null lock for every possible query type and as many
> exclusive sub-lock as needed, one for every possible instance of that
> particular query. for each new request of that type, the newly created
> object instance enqueue one exclusive lock on every sub-lock. as soon
> as one is granted, it first dequeue the unneeded ones and starts its
> job.
> 
> do you see any flow in this algorithm ? do you see any other way of
> doing it ?
> 
> TIA,
> Pierre.