Re: Impact of Increasing M_MMAP_THRESHOLD to 1GB in malloc().

[Carlos O'Donell <carlos at redhat dot com>]

On 9/27/18 1:45 PM, Florian Weimer wrote:
> * ritesh sonawane:
> 
>> Yes it is 64 Bit target.
>>
>> Fragmentation means when malloc() request is more than threshold
>> value, memory is allocated using mmap().  Due to size alignment with
>> page size, there is memory wastage per request.  e.g. Our system is
>> having 48GB memory, then total Number of malloc() requests (each
>> 64MB) will be 380 and total used memory is 23 GB out of 48GB.
> 
> I see.  I agree that's a problem, and changing the malloc threshold
> could be a solution.
> 
>> The system (NEC SX-Aurora TSUBASA) on which we are currently working
>> is having huge page size and each process can have 16 GB and 512 GB
>> address space for 2 MB and 64 MB page size respectively.
> 
> I'm not familiar with that system and haven't seen the glibc port,
> sorry.  I used to work next door to a NEC SX-6 as a student, but
> that's it.
> 
>> Also there is no particular limit on maximum heap size and Each
>> thread can easily consumes 2 GB of address space and thats why we
>> want to increase the M_MMAP_THRESHOLD to 1 GB.
> 
> If you do that, you also have to increase the heap size to something
> like 32 GiB (HEAP_MAX_SIZE in malloc/arena.c).  The default of 2 *
> DEFAULT_MMAP_THRESHOLD_MAX is probably too small (assuming that
> DEFAULT_MMAP_THRESHOLD_MAX will be 2 GiB).  Otherwise you will have
> substantial fragmentation for allocation requests between 2 GiB and
> HEAP_MAX_SIZE.

Right.

* Change malloc/malloc.c DEFAULT_MMAP_TRESHOLD_MAX to 16GiB.
* This in turn sets HEAP_MAX_SIZE to 32GiB.
* Now you can set M_MMAP_THRESHOLD to any value 0 > x <= (32GiB - a couple of pages).

The key here is that an arena is made up of discontiguous heaps in
a chain which are logically "the heap" for the attached thread (non-main
arena). You can't have a request that crosses heap boundaries, so if
you set M_MMAP_THRESHOLD larger than the heap size it is impossible to
service the request except through mmap.

All of this requires a custom glibc.

I *had* some patches to set HEAP_MAX_SIZE for testing, but it can only
be set at startup via a tunable because once set all allocations have
to use it for masking to compute chunk->heap mapping.

I never had a reason to change it... but this example is such a reason.

I don't think it's a terrible idea to allow a tunable for HEAP_MAX_SIZE
or DEFAULT_MMAP_THRESHOLD_MAX, but can only be set early and not late,
which is fine.

> I expect each heap will only allocate two pages via page faults, but
> will reserve HEAP_MAX_SIZE bytes of address space.

Agreed.

> If that's a problem, you could also make these changes, but set the
> maximum arena count to 1, then only the main arena will be used, which
> is sbrk-based.  The main arena doesn't need the coarse-grained
> mappings with large power-of-two sizes.

Right.

-- 
Cheers,
Carlos.