Re: Sv: Sv: Limit for number of child processes

[Ken Brown <kbrown@cornell.edu>]

On 8/28/2020 4:38 AM, sten.kristian.ivarsson@gmail.com wrote:
>>> Hi Corinna
>>>
>>>>> Dear cygwin folks
>>>>>
>>>>> It seems like there's a limit of the number of possible child
>>>>> processes defined to 256 with 'NPROCS' in
>>>>> //winsup/cygwin/child_info.h used in 'cprocs' in
>>>>> //winsup/cygwin/sigproc.cc
>>>>>
>>>>> 256 is quite few possible children in an enterprise environment and
>>>>> perhaps the limit should be limited by the physical resources or
>>>> possibly Windows ?
>>>>
>>>> The info has to be kept available in the process itself so we need
>>>> this array of NPROCS * sizeof (pinfo).
>>>>
>>>> Of course, there's no reason to use a static array, the code could
>>>> just as well use a dynamically allocated array or a linked list.
>>>> It's just not the way it is right now and would need a patch or
>> rewrite.
>>>>
>>>> As for the static array, sizeof pinfo is 64, so the current size of
>>>> the array is just 16K.  We could easily bump it to 64K with NPROCS
>>>> raised to
>>>> 1024 for the next Cygwin release, at least on 64 bit.
>>>> I don't think we should raise this limit for 32 bit Cygwin, which is
>>>> kind of EOL anyway, given the massive restrictions.
>>>
>>> I don't know the exact purpose of this and how the cprocs is used, but
>>> I'd prefer something totally dynamic 7 days out of 7 or otherwise
>>> another limit would just bite you in the ass some other day instead
>>> ;-)
>>>
>>> A linked list could be used if you wanna optimize (dynamic) memory
>>> usage but an (amortized) array would probably provide faster linear
>>> search but I guess simplicity of the code and external functionality
>>> is the most important demands for this choice
>>
>> Any change here (aside from just increasing NPROCS) would have to be done
>> with care to avoid a performance hit.  I looked at the history of changes
>> to sigproc.cc, and I found commit 4ce15a49 in 2001 in which a static array
>> something like cprocs was replaced by a dynamically allocated buffer in
>> order to save DLL space.  This was reverted 3 days later (commit e2ea684e)
>> because of performance issues.
> 
> 
> I wonder what kind of performance issue ? Nevertheless, that old commit
> didn't make the number of possible children more dynamic though, when it was
> still restricted to NPROCS (or ZOMBIEMAX/NZOMBIES), it was just not
> allocated on the stack. But yes, accessing dynamic allocated memory can
> theoretically be slower than stack allocated memory, but without measuring
> it one cannot tell ;-) Todays hardware is pretty good at prefetching etc,
> but as I said, it needs measurements

I don't know for sure, but I doubt if it had anything to do with memory access. 
My guess is that the performance hit came from the need to free the allocated 
memory after every fork call (see sigproc_fixup_after_fork).

Ken