Re: Pre-start Spawning jobs

[Tim Bronski <tim.bronski@xxxxxxxxx>]

With the multiple accept approach who is controlling events? How would 
you do a controlled shutdown or deal with varying loads? What happens 
when a child abends?


On 3/9/2017 4:22 PM, Chris Bipes wrote:
> Thank you for the outline.  I was reading in the IBM i: Sockets Programming guide where you can use multiple accept() APIs to handle incoming request.  I think I am going to go with that approach.  Do you know of any gotcha's with this method?
>
> --
> Chris Bipes
> Director of Information Services
> CrossCheck, Inc.
>
> -----Original Message-----
> From: MIDRANGE-L [mailto:midrange-l-bounces@xxxxxxxxxxxx] On Behalf Of Tim Bronski
> Sent: Thursday, March 09, 2017 2:24 AM
> To: Midrange Systems Technical Discussion <midrange-l@xxxxxxxxxxxx>
> Subject: Re: Pre-start Spawning jobs
>
> Hi Chris, I can tell you the outline of what you need to do if that
> helps. We have many server products that do this.
>
> Yes you will spawn off the child processes before you get a connection.
> We use a socket pair and sendmsg and recvmsg to communicate between the
> parent and children. The socket pair are mapped to the child job's file
> descriptors. We map to 0 and 1 and set this env var -
> "QIBM_USE_DESCRIPTOR_STDIO=Y". We send connection information to the
> children via sendmsg using a msghdr structure with the connection
> socket. You will need to make sure you catch child-has-ended signals so
> you can reset your counters otherwise you could end up with no child
> processes if they abend! Make sure you use waitpid with WNOHANG  after a
> child dies or you'll end up with zombie processes. We poll for acks from
> a child after we pass a connection to ensure we have active and ready
> children. The ack comes back on the other half of the sock pair that's
> set to non blocking. If we don't get an ack we verify the # of child
> jobs and possibly start another one to deal with volume. We do have a
> child number cap. If there are more than the initial number of child
> jobs we have an algorithm based on the last started child time that
> signals a child to end.