Experimental Physics and
Industrial Control System

Andrew Johnson <[email protected]> · Tue, 26 Apr 2011 10:19:09 -0500

On Tuesday 26 April 2011 08:06:52 Martin L. Smith wrote:
> If your gateways are running under Linux my understanding is that you may
>  want to increase the stack size limit. I have a comment in my starter
>  file:
> 
> # Limit stacksize for Linux to increase number of threads
> # 1024 is 204 threads (default)
> #  4096 is 512 threads
> ulimit -s 4096
> 
> As you can see I use 4096 although I'm not sure that this will help
>  especially if you are not running on Linux.

I want to add a little more information about what this actually does for 
future reference: Like most operating systems Linux has to reserve a fixed 
amount of virtual memory space for the program stack for each independent 
thread that gets started.  You can't move or extend a stack after the thread 
has started as each thread's stack is separated by the reserved distance from 
the next thread.

It depends on your particular distribution but that distance often defaults to 
around 10240KB=10MB.  If you're running on a 32-bit Intel CPU the OS usually 
only allows a process to use half of its virtual memory space, i.e. 2GB.  
2GB=2048MB divided by 10MB means that a single process such as the Gateway 
can't have more than 204 threads by default.  The CA client library starts two 
threads for each IOC it talks to, so that would limit the gateway or 
nameserver to connecting with at most about 100 IOCs.

Marty's "ulimit -s 4096" command sets the default stack size to 4096KB=4MB, 
thus permitting up to about 512 threads or 256 IOCs.  You can find the limit 
for a particular machine by running "ulimit -s" under Bash or "limit" under 
the C-shell.  In practice it should be possible to run the gateway code with a 
significantly smaller limit, but I don't know if anyone has tried to find out 
how small you can go — I would expect it to still work in 64K, although it 
seems unlikely that anyone would need a single gateway able to connect to 
32768 IOCs.

An alternative is to run the gateway on a 64-bit machine where the available 
virtual memory space is much larger, but I don't know whether the gateway code 
is 64-bit safe — has anyone tried that?

- Andrew
-- 
An error is only a mistake if you don't learn from it.
When you learn something from it, it becomes a lesson.

Experimental Physics and Industrial Control System

Experimental Physics and
Industrial Control System