hibernate and proc_lib

[Bjorn Gustavsson]

We have added the suggested hibernate/3 to proc_lib. It will be released
as a patch and also in R9C-2.

/Bjorn

"Ulf Wiger (AL/EAB)" <ulf.wiger@REDACTED> writes:

> There is a function called erlang:hibernate/3, which reduces the memory
> footprint of a process as much as possible until it receives a message.
> This function can be useful for applications that require a very large
> number of processes, but run into problems with memory consumption.
> Each time a process can be expected to idle for a while, one can call
> erlang:hibernate(Mod,Function,Args), and the process will be shrunk
> until it has reason to wake up, in which case apply(Mod,Function,Args)
> is called, and execution continues from there.
> 
> One of the things hibernate/3 does is remove the stack, which also 
> means that any 'catch' will also be removed. This is not a big problem, 
> as long as you're aware of it.
> 
> For example, if one has a top-level catch for central error handling, 
> this catch has to be re-inserted each time after hibernation. This 
> is reasonably easy to do if a wrapper function is provided. 
> 
> As an example, consider proc_lib:spawn_link(M,F,A).
> It is used e.g. by gen_servers and supervisors, and makes it possible to 
> receive crash reports if the process dies. This is accomplished using a 
> top-level catch in proc_lib:
> 
> init_p(Parent,Ancestors,M,F,A) ->
>     put('$ancestors',[Parent|Ancestors]),
>     put('$initial_call',{M,F,A}),
>     Result = (catch apply(M,F,A)),
>     exit_p(Result,{M,F,A}).
> 
> exit_p({'EXIT',Reason},StartF) ->
>     crash_report(Reason,StartF),
>     exit(Reason);
> exit_p(Reason,_) ->
>     Reason.
> 
> Doing hibernate on such a process would have the effect that
> no crash report would be generated after a crash, unless one
> provides a proc_lib:hibernate(M,F,A):
> 
> hibernate(M,F,A) ->
>    erlang:hibernate(proc_lib, wake_up, [M,F,A]).
> 
> wake_up(M,F,A) ->
>    Result = (catch apply(M,F,A)),
>    exit_p(Result, {M,F,A}).
> 
> I propose that OTP add such a function to proc_lib.erl in order
> to highlight the issue and demonstrate the workaround.
> 
> /Uffe
> 

-- 
Björn Gustavsson, Erlang/OTP, Ericsson AB