[v2] Add malloc micro benchmark

  Carlos O'Donell wrote:

> If you have a pattern of malloc/free of *similar* sized blocks, then
> it overflows the sized bin in the tcache, with other size bins remaining
> empty. The cache itself does not dynamically reconfigure itself to consume
> X MiB or Y % of RSS, instead it uses a simple data structure to contain
> a fixed number of fixed size blocks.
>
> Therefore I agree, that enhancing the core data structure in tcache may
> result in better overall performance, particularly if we got rid of the
> fixed bin sizes and instead found a way to be performant *and* keep a
> running total of consumption.

Well it could keep track of sum of all block sizes and limit that. That would
be better than a fixed limit on number of blocks in each bin. 

> Likewise *all* of malloc needs to be moved to a better data structure than
> just linked lists. I would like to see glibc's malloc offer a cacheing
> footprint of no more than Y % of RSS available, and let the user tweak that.
> Currently we just consume RSS without much regard for overhead. Though this
> is a different case than than what you are talking about, the changes are
> related via data-structure enhancements that would benefit both cases IMO.

What kind of datastructure do you have in mind? Small blocks could be
allocated together in pages. This would avoid the per-block overhead and
changes the linked list into a bitmap scan. However there aren't that many
alternatives. I've written first-fit allocators using autobalancing trees, but
walking the tree is expensive due to being not having good locality.

>> I *wish* we could test main_arena vs. threaded arena, since they
>> have different code and behave differently e.g. sbrk vs. mmap'd
>> heap.

I've added test of main and thread arenas in the latest version as well
as a test with a larger tcache.

> As I suggested in bug 15321:
> https://sourceware.org/bugzilla/show_bug.cgi?id=15321
>
> We need to merge the main_arena and threaded code together, and stop
> treating them as different things. Right now the main_arena, if you
> look at the code, is a *pretend* heap with a partial data structure
> layered in place. This needs to go away. We need to treat all heaps
> as identical, with identical code paths, with just different backing
> storage.

Yes that sounds like a good plan.

> I think people still expect that thread 0 allocates from the sbrk
> heap in a single-threaded application, and we can do that by ensuring
> sbrk is used to provide the backing store for the main thread. This way
> we can jump the pointer 64MB like we normally do for mmap'd heaps, but
> then on page touch there the kernel just extends the heap normally.

A quick benchmark shows sbrk is about ~25% faster than mmap, so it
appears useful. Looking at the addresses, sbrk grows up, while mmap
grows down the address space. I think other arenas could use sbrk as
well as long as you can free sbrk memory via MADV_FREE or
MADV_DONTUSE.


>> Implementation:
>>
>> You need to make this robust against env vars changing malloc
>> behaviour. You should use mallopt to change some parameters.

I've added support to change the tcache count in mallopt, so we can
benchmark different settings.

> We need to move to MADV_FREE, which was designed for memory allocators.
>
> The semantics of MADV_DONTNEED have the problem that one has to consider:
> * Is the data destructively lost in that page?
> * Is the data flushed to the underlying store before being not-needed?
> All of which lead to MADV_DONTNEED doing a lot of teardown work to ensure
> that users don't corrupt the data in their backing stores.
>
> I think that detection of MADV_FREE, and usage, would help performance,
> but only on > Linux 4.5, and that might be OK for you.

Well we should detect when MADV_FREE is supported and use that.
If not, tweak the size of MADV_DONTNEED - perhaps based on a
percentage of the RSS size so we limit the overhead.


Anyway, here is the new version:


Add a malloc micro benchmark to enable accurate testing of the
various paths in malloc and free.  The benchmark does a varying
number of allocations of a given block size, then frees them again.

It tests 4 different scenarios: single-threaded using main arena,
multi-threaded using thread-arena, main arena with SINGLE_THREAD_P
false and main arena with the tcache count set larger than the
default.  To enable this, add support for M_TCACHE_COUNT in mallopt.

OK for commit?

ChangeLog:
2018-01-02  Wilco Dijkstra  <wdijkstr@arm.com>

	* benchtests/Makefile: Add malloc-simple benchmark.
	* benchtests/bench-malloc-simple.c: New benchmark.
	* malloc/malloc.h (M_TCACHE_COUNT): Add new define.
	* malloc/malloc.c (__libc_mallopt): Handle M_TCACHE_COUNT.

--
  

DJ Delorie wrote:

> What other tests do is create a second test that just #include's the
> first test, and set an environment variable in the Makefile specific to
> that test.  Adding an ABI just for a test is a big hammer, although we
> could discuss adding tcache to mallopt() as a seperate topic.

Yeah but the makefiles are already insanely complex. Adding the new
benchmark to the makefile took more than 10x as much time as writing
the test itself...

> I don't have any objection to adding tcache to mallopt (although please
> add all three tunables if you do), just saying we should discuss it as
> an ABI change separately.

It doesn't have to be an external ABI, I'd suggest keeping this internal to
GLIBC to make testing and benchmarking easier.

Wilco
  

On 01/03/2018 04:12 AM, Wilco Dijkstra wrote:
> DJ Delorie wrote:
> 
>> What other tests do is create a second test that just #include's the
>> first test, and set an environment variable in the Makefile specific to
>> that test.  Adding an ABI just for a test is a big hammer, although we
>> could discuss adding tcache to mallopt() as a seperate topic.
> 
> Yeah but the makefiles are already insanely complex. Adding the new
> benchmark to the makefile took more than 10x as much time as writing
> the test itself...
> 
>> I don't have any objection to adding tcache to mallopt (although please
>> add all three tunables if you do), just saying we should discuss it as
>> an ABI change separately.
> 
> It doesn't have to be an external ABI, I'd suggest keeping this internal to
> GLIBC to make testing and benchmarking easier.

Don't use mallopt, please make it a tunable then.

The mallopt API already had 2 secret arena options which eventually became
so well used they were baked into the API and had to be made public.

At least with tunables we are allowed to deprecate them.
  

Carlos O'Donell wrote:
>
> Don't use mallopt, please make it a tunable then.
> 
> The mallopt API already had 2 secret arena options which eventually became
> so well used they were baked into the API and had to be made public.

Unfortunately tunables are not exported so you can't use them outside of GLIBC:

/build/glibc/benchtests/bench-malloc-simple.o: In function `bench':
bench-malloc-simple.c:(.text+0x19c): undefined reference to `__tunable_set_val'
collect2: error: ld returned 1 exit status

Wilco
  

On 04/01/2018 11:48, Wilco Dijkstra wrote:
> Carlos O'Donell wrote:
>>
>> Don't use mallopt, please make it a tunable then.
>>
>> The mallopt API already had 2 secret arena options which eventually became
>> so well used they were baked into the API and had to be made public.
> 
> Unfortunately tunables are not exported so you can't use them outside of GLIBC:
> 
> /build/glibc/benchtests/bench-malloc-simple.o: In function `bench':
> bench-malloc-simple.c:(.text+0x19c): undefined reference to `__tunable_set_val'
> collect2: error: ld returned 1 exit status
> 
> Wilco
> 

You will need to the environment variable to check outside GLIBC (as expected
by normal programs).
  

On 01/04/2018 05:48 AM, Wilco Dijkstra wrote:
> Carlos O'Donell wrote:
>>
>> Don't use mallopt, please make it a tunable then.
>>
>> The mallopt API already had 2 secret arena options which eventually became
>> so well used they were baked into the API and had to be made public.
> 
> Unfortunately tunables are not exported so you can't use them outside of GLIBC:
> 
> /build/glibc/benchtests/bench-malloc-simple.o: In function `bench':
> bench-malloc-simple.c:(.text+0x19c): undefined reference to `__tunable_set_val'
> collect2: error: ld returned 1 exit status

Correct, we only have a env-var frontend right now, and the internal API is not
made accessible via GLIBC_PRIVATE.

You have 3 options for tests:

* Use the env vars to adjust test behaviour. Run the tests multiple times.
* Add a new C API frontend, very valuable, but more time consuming.
* Expose the existing internal C API via GLIBC_PRIVATE for testing, and throw
  it away later when we get a proper C API frontend.
  

On 01/02/2018 10:20 AM, Wilco Dijkstra wrote:
> Carlos O'Donell wrote:
> 
>> If you have a pattern of malloc/free of *similar* sized blocks, then
>> it overflows the sized bin in the tcache, with other size bins remaining
>> empty. The cache itself does not dynamically reconfigure itself to consume
>> X MiB or Y % of RSS, instead it uses a simple data structure to contain
>> a fixed number of fixed size blocks.
>>
>> Therefore I agree, that enhancing the core data structure in tcache may
>> result in better overall performance, particularly if we got rid of the
>> fixed bin sizes and instead found a way to be performant *and* keep a
>> running total of consumption.
> 
> Well it could keep track of sum of all block sizes and limit that. That would
> be better than a fixed limit on number of blocks in each bin. 

Yes, in practice what we want to enforce is % of total RSS, or fixed X MiB of
RSS in thread caches, and we want that value to be a per-thread variable that can
be changed for each thread depending on the thread's workload.

e.g.

Thread 1, 2, and 3 each need 5% of RSS for their workloads.

Threads 4, and 5, need 25% of RSS for their workloads.

Obviously, if RSS is very low, then our 32MiB/64MiB starting heaps may be unusable,
but could also be tuned, it's a decision that can only be made once at startup
because we use this embedded assumption for pointer arithmetic to find the arena
structure address. It could be changed, but this would increase costs.

>> Likewise *all* of malloc needs to be moved to a better data structure than
>> just linked lists. I would like to see glibc's malloc offer a cacheing
>> footprint of no more than Y % of RSS available, and let the user tweak that.
>> Currently we just consume RSS without much regard for overhead. Though this
>> is a different case than than what you are talking about, the changes are
>> related via data-structure enhancements that would benefit both cases IMO.
> 
> What kind of datastructure do you have in mind? Small blocks could be
> allocated together in pages. This would avoid the per-block overhead and
> changes the linked list into a bitmap scan. However there aren't that many
> alternatives. I've written first-fit allocators using autobalancing trees, but
> walking the tree is expensive due to being not having good locality.

Keep in mind we are deviating now from the topic at hand, but I'll lay out two
improvements, one of which you already touched upon.

(1) Small blocks cost too much RSS.

I have had customers hit terrible corner cases with C++ and small blocks which
see ~50% waste due to colocated metadata e.g. new of 13-byte objects.

Getting smaller allocations working together for the small blocks would be a big win,
and using some kind of bitmap would be my suggested solution. This requires a data
structure to track the parent allocation, bitmaps, and sub-allocations. We have some
of these data structures in place, but no easy generic way to use them (I think Florian's
pushing of more general data structure use in glibc is the right way to go e.g. dynarray).

I think that for blocks smaller than the fundamental language types (which require
malloc to have 16-byte alignment) we do not have to return sufficiently aligned
memory. For example if you allocate a 3-byte block or a 13-byte block, you cannot
possibly put a 16-byte long double there, nor can you use that for a stack block,
so it's a waste to guarantee alignment.

* Group small allocations.
* Violate the ABI and use < MALLOC_ALIGNMENT sized alignments for sub-group members.

(2) Track total memory used and free back based on more consistent heuristics.

Again, we have customers who complain that glibc's malloc is bad for container
or VM workloads with tight packing because it just keeps allocating more heaps.

We could do better to track free page runs, and when we exceed some %RSS
threshold, free back larger runs to the OS, something which malloc_trim()
already does but in a more costly manner by walking the whole arena heaps from
start to end.

This isn't explicitly about performance.

(3) Make arena's truly a per-cpu data structure.

We do not want per-thread arenas.

We want per-thread caches (avoids locks)

We want per-cpu arenas (avoid NUMA issues, and get better locality)

The per-thread cache does the job of caching the thread-local requests
and avoiding locks.

The per-cpu arenas do the job of containing cpu-local memory, and handling
requests for the threads that reside on that CPU, either pinned, or there
temporarily.

Memory being returned should go back to the cpu that the memory came from.

The best we have done today is to have the arenas scale with the number of
CPUs, and let them fall where they may across the cores, and let the
numa page scheduler move the memory around them to minimize cost.

The best way would be to use something like restartable sequences to
get the CPU #, select the arena, and get memory from there, and then
choose another arena next time perhaps.

(4) Distribute threads among arenas.

We have no data structure for balancing arena usage across threads.

We have seen workloads where the maximum number of arenas is allocated,
but threads don't move smoothly across arenas, instead the first unlocked
arena is chosen, and that might not be the best from a memory locality
perspective (see (3)).

> Add a malloc micro benchmark to enable accurate testing of the
> various paths in malloc and free.  The benchmark does a varying
> number of allocations of a given block size, then frees them again.
> 
> It tests 4 different scenarios: single-threaded using main arena,
> multi-threaded using thread-arena, main arena with SINGLE_THREAD_P
> false and main arena with the tcache count set larger than the
> default.  To enable this, add support for M_TCACHE_COUNT in mallopt.
> 
> OK for commit?
> 
> ChangeLog:
> 2018-01-02  Wilco Dijkstra  <wdijkstr@arm.com>
> 
> 	* benchtests/Makefile: Add malloc-simple benchmark.
> 	* benchtests/bench-malloc-simple.c: New benchmark.
> 	* malloc/malloc.h (M_TCACHE_COUNT): Add new define.
> 	* malloc/malloc.c (__libc_mallopt): Handle M_TCACHE_COUNT.

I don't like the creation of a new public ABI via M_TCACHE_COUNT through mallopt.

I suggest splitting these tests apart and using the tunable env var to touch this.

See my other email.
  

On 05/01/2018 12:32, Carlos O'Donell wrote:
> On 01/04/2018 05:48 AM, Wilco Dijkstra wrote:
>> Carlos O'Donell wrote:
>>>
>>> Don't use mallopt, please make it a tunable then.
>>>
>>> The mallopt API already had 2 secret arena options which eventually became
>>> so well used they were baked into the API and had to be made public.
>>
>> Unfortunately tunables are not exported so you can't use them outside of GLIBC:
>>
>> /build/glibc/benchtests/bench-malloc-simple.o: In function `bench':
>> bench-malloc-simple.c:(.text+0x19c): undefined reference to `__tunable_set_val'
>> collect2: error: ld returned 1 exit status
> 
> Correct, we only have a env-var frontend right now, and the internal API is not
> made accessible via GLIBC_PRIVATE.
> 
> You have 3 options for tests:
> 
> * Use the env vars to adjust test behaviour. Run the tests multiple times.
> * Add a new C API frontend, very valuable, but more time consuming.
> * Expose the existing internal C API via GLIBC_PRIVATE for testing, and throw
>   it away later when we get a proper C API frontend.
> 

Do we want a C API to tied the malloc implementation to some tunables? My
understanding is the tunable api idea is not really enforce retro-compability
(where a C api would enforce it).
  

On 01/05/2018 07:50 AM, Adhemerval Zanella wrote:
> 
> 
> On 05/01/2018 12:32, Carlos O'Donell wrote:
>> On 01/04/2018 05:48 AM, Wilco Dijkstra wrote:
>>> Carlos O'Donell wrote:
>>>>
>>>> Don't use mallopt, please make it a tunable then.
>>>>
>>>> The mallopt API already had 2 secret arena options which eventually became
>>>> so well used they were baked into the API and had to be made public.
>>>
>>> Unfortunately tunables are not exported so you can't use them outside of GLIBC:
>>>
>>> /build/glibc/benchtests/bench-malloc-simple.o: In function `bench':
>>> bench-malloc-simple.c:(.text+0x19c): undefined reference to `__tunable_set_val'
>>> collect2: error: ld returned 1 exit status
>>
>> Correct, we only have a env-var frontend right now, and the internal API is not
>> made accessible via GLIBC_PRIVATE.
>>
>> You have 3 options for tests:
>>
>> * Use the env vars to adjust test behaviour. Run the tests multiple times.
>> * Add a new C API frontend, very valuable, but more time consuming.
>> * Expose the existing internal C API via GLIBC_PRIVATE for testing, and throw
>>   it away later when we get a proper C API frontend.
>>
> 
> Do we want a C API to tied the malloc implementation to some tunables? My
> understanding is the tunable api idea is not really enforce retro-compability
> (where a C api would enforce it).
 
If we add a C API to the tunables, we would honour that API for tunables for
all time, but the tunables themselves would not be stable.

e.g.

* get list of tunables supported
* get the default value for a tunable
* get the value of a tunable
* set the value of a tunable

So you would use this API in the tests to get the tunable list, assert the
tcache tunable was accepted (or fail the test), and then set it to a special
value for the part of the test that needs it.
  

On Fri, 5 Jan 2018, Carlos O'Donell wrote:

> I think that for blocks smaller than the fundamental language types 
> (which require malloc to have 16-byte alignment) we do not have to 
> return sufficiently aligned memory. For example if you allocate a 3-byte 
> block or a 13-byte block, you cannot possibly put a 16-byte long double 
> there, nor can you use that for a stack block, so it's a waste to 
> guarantee alignment.

As per DR#075, the memory needs to be aligned for any type of object (with 
a fundamental alignment requirement, in C11 and later), not just those 
that will fit in the block.  (This in turn allows for applications using 
low bits for tagged pointers.)

This does not of course rule out having another allocation API that 
supports smaller alignment requirements.
  

On 05/01/2018 14:17, Carlos O'Donell wrote:
> On 01/05/2018 07:50 AM, Adhemerval Zanella wrote:
>>
>>
>> On 05/01/2018 12:32, Carlos O'Donell wrote:
>>> On 01/04/2018 05:48 AM, Wilco Dijkstra wrote:
>>>> Carlos O'Donell wrote:
>>>>>
>>>>> Don't use mallopt, please make it a tunable then.
>>>>>
>>>>> The mallopt API already had 2 secret arena options which eventually became
>>>>> so well used they were baked into the API and had to be made public.
>>>>
>>>> Unfortunately tunables are not exported so you can't use them outside of GLIBC:
>>>>
>>>> /build/glibc/benchtests/bench-malloc-simple.o: In function `bench':
>>>> bench-malloc-simple.c:(.text+0x19c): undefined reference to `__tunable_set_val'
>>>> collect2: error: ld returned 1 exit status
>>>
>>> Correct, we only have a env-var frontend right now, and the internal API is not
>>> made accessible via GLIBC_PRIVATE.
>>>
>>> You have 3 options for tests:
>>>
>>> * Use the env vars to adjust test behaviour. Run the tests multiple times.
>>> * Add a new C API frontend, very valuable, but more time consuming.
>>> * Expose the existing internal C API via GLIBC_PRIVATE for testing, and throw
>>>   it away later when we get a proper C API frontend.
>>>
>>
>> Do we want a C API to tied the malloc implementation to some tunables? My
>> understanding is the tunable api idea is not really enforce retro-compability
>> (where a C api would enforce it).
>  
> If we add a C API to the tunables, we would honour that API for tunables for
> all time, but the tunables themselves would not be stable.
> 
> e.g.
> 
> * get list of tunables supported
> * get the default value for a tunable
> * get the value of a tunable
> * set the value of a tunable
> 
> So you would use this API in the tests to get the tunable list, assert the
> tcache tunable was accepted (or fail the test), and then set it to a special
> value for the part of the test that needs it.

Right, this seems a reasonable approach (although I think out of the scope for
this change).
  

On 01/05/2018 08:28 AM, Joseph Myers wrote:
> On Fri, 5 Jan 2018, Carlos O'Donell wrote:
> 
>> I think that for blocks smaller than the fundamental language types 
>> (which require malloc to have 16-byte alignment) we do not have to 
>> return sufficiently aligned memory. For example if you allocate a 3-byte 
>> block or a 13-byte block, you cannot possibly put a 16-byte long double 
>> there, nor can you use that for a stack block, so it's a waste to 
>> guarantee alignment.
> 
> As per DR#075, the memory needs to be aligned for any type of object (with 
> a fundamental alignment requirement, in C11 and later), not just those 
> that will fit in the block.  (This in turn allows for applications using 
> low bits for tagged pointers.)

Thanks for the reference to DR#075, I had not considered the cast equality
issue.

> This does not of course rule out having another allocation API that 
> supports smaller alignment requirements.
Agreed.

It would still be a win if we did not have co-located metadata (something
Florian whispered into my ear years ago now) for small constant sized blocks.

We would go from this:

N * 1-byte allocations => N * (32-byte header 
                               + 1-byte allocation
                               + 15-bytes alignment)
                          [97% constant waste]

To this:

N * 1-byte allocations => N * (1-byte allocation
                               + 15-bytes alignment) 
                          + (N/8)-bytes in-use-bit + 16-bytes header
                          [96% waste for 1-byte]
			  [94% waste for 100*1-byte]
                          ... towards a 93.75% constant waste (limit of the alignment e.g. 15/16)

This is a gain of 5% RSS efficiency for a structural change.

For a 13-byte allocation:

N * 1-byte allocations => N * (32-byte header 
                               + 13-byte allocation
                               + 3-bytes alignment)
                          [73% constant waste]

To this:

N * 1-byte allocations => N * (13-byte allocation
                               + 3-bytes alignment) 
                          + (N/8)-bytes in-use-bit + 16-bytes header
                          [60% waste for 13-bytes]
			  [20% waste for 100*13-bytes]
			  [19% waste for 1000*13-bytes]
			  ... towards a 18.75% constant waste (limit of the alignment e.g. 3/16)

Note: We never reach the constant limit because the in-use bit-array still grows quickly.
  

On 01/05/2018 08:46 AM, Adhemerval Zanella wrote:
> 
> 
> On 05/01/2018 14:17, Carlos O'Donell wrote:
>> On 01/05/2018 07:50 AM, Adhemerval Zanella wrote:
>>>
>>>
>>> On 05/01/2018 12:32, Carlos O'Donell wrote:
>>>> On 01/04/2018 05:48 AM, Wilco Dijkstra wrote:
>>>>> Carlos O'Donell wrote:
>>>>>>
>>>>>> Don't use mallopt, please make it a tunable then.
>>>>>>
>>>>>> The mallopt API already had 2 secret arena options which eventually became
>>>>>> so well used they were baked into the API and had to be made public.
>>>>>
>>>>> Unfortunately tunables are not exported so you can't use them outside of GLIBC:
>>>>>
>>>>> /build/glibc/benchtests/bench-malloc-simple.o: In function `bench':
>>>>> bench-malloc-simple.c:(.text+0x19c): undefined reference to `__tunable_set_val'
>>>>> collect2: error: ld returned 1 exit status
>>>>
>>>> Correct, we only have a env-var frontend right now, and the internal API is not
>>>> made accessible via GLIBC_PRIVATE.
>>>>
>>>> You have 3 options for tests:
>>>>
>>>> * Use the env vars to adjust test behaviour. Run the tests multiple times.
>>>> * Add a new C API frontend, very valuable, but more time consuming.
>>>> * Expose the existing internal C API via GLIBC_PRIVATE for testing, and throw
>>>>   it away later when we get a proper C API frontend.
>>>>
>>>
>>> Do we want a C API to tied the malloc implementation to some tunables? My
>>> understanding is the tunable api idea is not really enforce retro-compability
>>> (where a C api would enforce it).
>>  
>> If we add a C API to the tunables, we would honour that API for tunables for
>> all time, but the tunables themselves would not be stable.
>>
>> e.g.
>>
>> * get list of tunables supported
>> * get the default value for a tunable
>> * get the value of a tunable
>> * set the value of a tunable
>>
>> So you would use this API in the tests to get the tunable list, assert the
>> tcache tunable was accepted (or fail the test), and then set it to a special
>> value for the part of the test that needs it.
> 
> Right, this seems a reasonable approach (although I think out of the scope for
> this change).
 
That is up to Wilco to decide, but in general I agree that he need not take on
this work to get the current patch set merged, there are other solutions to the
need to tweak the settings. I think the env var and multiple test run approach
is going to be the simplest.
  

On 01/05/2018 06:26 PM, Carlos O'Donell wrote:
> It would still be a win if we did not have co-located metadata (something
> Florian whispered into my ear years ago now) for small constant sized blocks.
> 
> We would go from this:
> 
> N * 1-byte allocations => N * (32-byte header
>                                 + 1-byte allocation
>                                 + 15-bytes alignment)
>                            [97% constant waste]

Actually, we have an 8-byte header, a 16-byte alignment requirement, and 
a 24-byte minimum allocation size.  (i386: 4-byte header, 16-byte 
alignment, 12-byte minimum size.)

The non-main arenas actually need only a four-byte chunk header (or even 
fewer bits) because there, the chunk size is very limited.

The alignment is non-negotiable, considering our current stance 
regarding fundamental alignment, even for smaller allocations.

If we introduce heap layout for all arenas (which needs some way to 
compensate for the variable sbrk offset, preferably without wasting a 
gap there), then we should bring down the minimum allocation size to 12 
bytes on 64-bit as well because for the smallest bin, we can use 4-byte 
forward/backward links within a heap, and keep 8-byte pointers separate 
for each heap.

None of this needs algorithm changes, but it's still difficult to 
identify all the places which need changing, due to the code duplication 
and some other issues with the code.

But it only helps with oddly-sized allocations (such as 12 bytes or 28 
bytes), so it is unclear whether this work is worthwhile.

> To this:
> 
> N * 1-byte allocations => N * (1-byte allocation
>                                 + 15-bytes alignment)
>                            + (N/8)-bytes in-use-bit + 16-bytes header
>                            [96% waste for 1-byte]
> 			  [94% waste for 100*1-byte]
>                            ... towards a 93.75% constant waste (limit of the alignment e.g. 15/16)

Another significant win is for allocation sizes such as 32, where we 
currently need to allocate 48 bytes.  In fact, powers-of-two are quite 
bad for the current allocator.

However, it is difficult to combine this with the existing allocator 
because you need to perform some table lookup during free to discover 
whether the pointer has adjacent metadata or not, or waste some address 
space and use bits inside the address for that.

If we want to bring the existing allocator further along, we should 
perhaps try to port dlmalloc changes which get rid of the unsorted bins, 
using balanced binary trees.  I have a feeling that this would allow us 
to consolidate far more often, and this should help us to avoid some of 
the anomalies we have seen.

Thanks,
Florian
  

On Wed, Feb 28, 2018 at 01:40:28PM +0100, Florian Weimer wrote:
> On 01/05/2018 06:26 PM, Carlos O'Donell wrote:
> >It would still be a win if we did not have co-located metadata (something
> >Florian whispered into my ear years ago now) for small constant sized blocks.
> >
> >We would go from this:
> >
> >N * 1-byte allocations => N * (32-byte header
> >                                + 1-byte allocation
> >                                + 15-bytes alignment)
> >                           [97% constant waste]
> 
> Actually, we have an 8-byte header, a 16-byte alignment requirement,
> and a 24-byte minimum allocation size.  (i386: 4-byte header,
> 16-byte alignment, 12-byte minimum size.)
> 
> The non-main arenas actually need only a four-byte chunk header (or
> even fewer bits) because there, the chunk size is very limited.
> 
> The alignment is non-negotiable, considering our current stance
> regarding fundamental alignment, even for smaller allocations.
>
Thats rather ineffective, it is easier to start fresh than try to
maintain rather obsolete allocator. Most of other are faster and more
space effective because of their layout.

I have several ideas based on that all allocations on a page of size P must 
have same capacity(say upto 256 bytes), options where metadata for pointer x are due 
to possibilitity of bigger alignment requirements following

1. End of previous page, easiest to implement. formula P*(x/P)-16

2. Page map, on 64-bit systems its possible on mmap also allocate pages
for mapping 16*(x/P). On 32-bit its possible add constant and rezerve
pages to map entire address space with same formula.

3. Encode size to address to table lookup in free for say size<=256. This trick is
probably applicable only for 64-bit systems as it could cause problems
with virtual address space. First calculate size by formula 16*((x-start)/(1<<25))
if its less than 257 we got correct size, otherwise use option 1/2
  

On 02/28/2018 03:11 PM, Ondřej Bílka wrote:
> Thats rather ineffective, it is easier to start fresh than try to
> maintain rather obsolete allocator. Most of other are faster and more
> space effective because of their layout.

That's not quite true.  Despite its limitations, glibc malloc still 
compares remarkably well to other allocators.  Of course, there are 
workloads where it loses big, but those exist for other allocators, too. 
  People simple don't write blog posts comparing *alloc with glibc 
malloc if glibc malloc provides comparable or better performance because 
it's quite boring.

I think a heap-style allocator which does not segregate allocations of 
different sizes still has its place, and why not provide one in glibc?

Thanks,
Florian
  

On 02/28/2018 06:16 AM, Florian Weimer wrote:
> On 02/28/2018 03:11 PM, Ondřej Bílka wrote:
>> Thats rather ineffective, it is easier to start fresh than try to 
>> maintain rather obsolete allocator. Most of other are faster and
>> more space effective because of their layout.
> 
> That's not quite true.  Despite its limitations, glibc malloc still
> compares remarkably well to other allocators.  Of course, there are
> workloads where it loses big, but those exist for other allocators,
> too.  People simple don't write blog posts comparing *alloc with
> glibc malloc if glibc malloc provides comparable or better
> performance because it's quite boring.
> 
> I think a heap-style allocator which does not segregate allocations
> of different sizes still has its place, and why not provide one in
> glibc?

I agree.

I think an incremental improvement would be to start with some further
code cleanups, all with the goal of simplifying the allocator maintenance.

Getting rid of unsorted bins would be interesting. One of the problems I
see is that without changes in our data structures and algorithms it is hard
for a heap-style allocator to keep a bound on the total amount of heap consumed.

One of the big requirements that I see coming down the pipe is to provide an
allocator that has bounded memory usage. To do that we'd have to track all of
the unsorted blocks, free blocks, consolidate more regularly, and free against
the heuristic which tells us how much memory to keep cached. If having got rid
of unsorted makes this easier, if having a balanced binary tree makes this
easier, then that's where we should be looking.
  

On Wed, Feb 28, 2018 at 03:16:09PM +0100, Florian Weimer wrote:
> On 02/28/2018 03:11 PM, Ondřej Bílka wrote:
> >Thats rather ineffective, it is easier to start fresh than try to
> >maintain rather obsolete allocator. Most of other are faster and more
> >space effective because of their layout.
> 
> That's not quite true.  Despite its limitations, glibc malloc still
> compares remarkably well to other allocators.

That confuses cause and effect. People spent lot of effort to fix
workloads where this allocator is bad (usually starts with saying that
malloc is slow) with trick like static buffers, merging allocations, memory pools etc. 
With better allocator it wouldn't be neccessary as it would handle simpler 
code with same performance. 
> 
> I think a heap-style allocator which does not segregate allocations
> of different sizes still has its place, and why not provide one in
> glibc?
>
That isn't case for any allocator and it is asking for trouble. You want
to avoid sitation where two big chunks couldn't be merged because of
tiny chunk between them. 
Also you will likely spend more space on header overhead than could you save.
For small sizes merging chunks doesn't happen so we lose nothing by
requiring uniform capacity. 

For larger size this representation is still problematic and you could
improve performance with another representation that also avoids
alignment problem by placing metadata elsewhere(for mine only 4 bytes are needed).
  

Ondřej Bílka wrote:
  
>> I think a heap-style allocator which does not segregate allocations
>> of different sizes still has its place, and why not provide one in
>> glibc?
>>
> That isn't case for any allocator and it is asking for trouble. You want
> to avoid sitation where two big chunks couldn't be merged because of
> tiny chunk between them.

Agreed, you always want to special case small blocks. I don't believe there is
any advantage in using a single big heap.

> For larger size this representation is still problematic and you could
> improve performance with another representation that also avoids
> alignment problem by placing metadata elsewhere(for mine only 4 bytes are needed).

Larger sizes would be helped a lot once small blocks are dealt with separately.
So I don't think we need complicated balanced binary trees when dealing with a
small number of large blocks. You won't need an unsorted list either, large blocks
can be merged in O(1) time.

There may be an advantage to place meta data elsewhere, for example it could make
adding/removing/walking free lists much faster (spatial locality) or to make heap
overflow attacks almost impossible,

Wilco
  

On 02/28/2018 09:01 AM, Wilco Dijkstra wrote:
> Ondřej Bílka wrote:
>   
>>> I think a heap-style allocator which does not segregate allocations
>>> of different sizes still has its place, and why not provide one in
>>> glibc?
>>>
>> That isn't case for any allocator and it is asking for trouble. You want
>> to avoid sitation where two big chunks couldn't be merged because of
>> tiny chunk between them.
> 
> Agreed, you always want to special case small blocks. I don't believe there is
> any advantage in using a single big heap.
> 
>> For larger size this representation is still problematic and you could
>> improve performance with another representation that also avoids
>> alignment problem by placing metadata elsewhere(for mine only 4 bytes are needed).
> 
> Larger sizes would be helped a lot once small blocks are dealt with separately.
> So I don't think we need complicated balanced binary trees when dealing with a
> small number of large blocks. You won't need an unsorted list either, large blocks
> can be merged in O(1) time.
> 
> There may be an advantage to place meta data elsewhere, for example it could make
> adding/removing/walking free lists much faster (spatial locality) or to make heap
> overflow attacks almost impossible,

I agree with many of the things you and Ondrej are proposing.

The outcome of our current discussions should be an incremental plan to go from
where we are today, to where we want to be tomorrow.

However, I do *not* believe it is a good plan to simply throw away the present
allocator and claim it should be replaced from scratch. We do not have that luxury
as a core project, we must remain answerable to our users.

A high-level concrete problem today with glibc's malloc, and the only problem being
reported by our users is that it consumes too much RSS. Solving that problem in the
abstract is what we should be looking at.

If we think that having multiple heaps for different sized objects is the way to
do this, then we should think about how to go down that path with an experiment.

Any cleanup we do before that is a win.
  

On Wed, Feb 28, 2018 at 05:01:57PM +0000, Wilco Dijkstra wrote:
> Ondřej Bílka wrote:
>   
> >> I think a heap-style allocator which does not segregate allocations
> >> of different sizes still has its place, and why not provide one in
> >> glibc?
> >>
> > That isn't case for any allocator and it is asking for trouble. You want
> > to avoid sitation where two big chunks couldn't be merged because of
> > tiny chunk between them.
> 
> Agreed, you always want to special case small blocks. I don't believe there is
> any advantage in using a single big heap.
> 
> > For larger size this representation is still problematic and you could
> > improve performance with another representation that also avoids
> > alignment problem by placing metadata elsewhere(for mine only 4 bytes are needed).
> 
> Larger sizes would be helped a lot once small blocks are dealt with separately.
> So I don't think we need complicated balanced binary trees when dealing with a
> small number of large blocks. You won't need an unsorted list either, large blocks
> can be merged in O(1) time.
> 
> There may be an advantage to place meta data elsewhere, for example it could make
> adding/removing/walking free lists much faster (spatial locality) or to make heap
> overflow attacks almost impossible,
>
I will answer now what I plan for larger blocks, 
I have new data structure mostly in head so I won't put concrete example. 

First for small sizes allocation would be just poping element from
thread local single linked list, or calling function to refill lists
with enough elements when empty. I plan to add inline version to make performance 
of constant small allocations same as memory pool. By using pointer to
list refill could do best-fit by making multiple buckets point to same
stack.
This is pretty generic interface, question is for which sizes it should
be used.

For larger I could do best fit in O(1) with merging on free. 
It needs a condition like that we are rounding up size/alignment 
to multiple of 32 for 256-2048 range and 256 for 2048-16384 as example. 
Data structure would be 64 double-linked lists and 64bit integer where 
i-th bit says if i-th list is nonempty. Last bucket could be special to
hold larger elements.

Finally for larger allocations I would use page-based logic as
mmaping/remapping/unmapping is about only way to actually decrease
memory footprint, I didn't try that much yet.

Code for allocation would be something like this

if (size < 256)
  {
    bucket = (size + 15) / 16;
    return small_list_pop (small_list[bucket]);
  }
else if (size < 32 * 64)
  {
    bucket = (size + 31) / 32
    uint64_t t = bitmap & ((-1) << bucket);
    if (t)
      bucket = __builtin_ctzl (t);
    else
      bucket = allocate (size);
    return list_pop(bucket);
  }
else if (size < 256 * 64)
    bucket = (size + 255) / 256;
  /* ditto with bigger buckets */
else
  /* mmap */


As free for small sizes I didn't decided yet how reclaim that to cache. 
For inlining it could be something simple like create single linked list of 32 elements, 
then call mass free for that list.

For medium elements it would first determine free areas before and after
free chunk, remove them from their double linked lists and unset bit if
necessary. Then sum these sizes and put it into appropriate bucket.
  

On Wed, Feb 28, 2018 at 08:16:13AM -0800, Carlos O'Donell wrote:
> On 02/28/2018 06:16 AM, Florian Weimer wrote:
> > On 02/28/2018 03:11 PM, Ondřej Bílka wrote:
> >> Thats rather ineffective, it is easier to start fresh than try to 
> >> maintain rather obsolete allocator. Most of other are faster and
> >> more space effective because of their layout.
> > 
> > That's not quite true.  Despite its limitations, glibc malloc still
> > compares remarkably well to other allocators.  Of course, there are
> > workloads where it loses big, but those exist for other allocators,
> > too.  People simple don't write blog posts comparing *alloc with
> > glibc malloc if glibc malloc provides comparable or better
> > performance because it's quite boring.
> > 
> > I think a heap-style allocator which does not segregate allocations
> > of different sizes still has its place, and why not provide one in
> > glibc?
> 
> I agree.
> 
> I think an incremental improvement would be to start with some further
> code cleanups, all with the goal of simplifying the allocator maintenance.
>
You should like I did try to decruft implementation. I decided that
starting again is simpler after looking lot on existing how to do it. I
send some patches with decrufting but I found that with changing
algorithm, data structures mmap logic and basically everything else its
just unnecessary overhead.

You couldn't decrease data structure overhead without changing data
structure.
And for RSS size problem is in design that you couldn't return memory to
system. Large areas get pined by small allocations with return data and
you couldn't sbrk. It is needed to redesign it to unmap pages
individually.
Alternative would be to add something like malloca with separate arenas
for which existing logic works.
  

On Wed, Feb 28, 2018 at 04:56:28PM -0500, DJ Delorie wrote:
> 
> Ond?ej B?lka <neleai@seznam.cz> writes:
> > First for small sizes allocation would be just popping element from
> > thread local single linked list, or calling function to refill lists
> > with enough elements when empty.
> 
> This is basically what tcache does.  Although I tested a few ways of
> pre-filling the cache, there's room for more research there beyond the
> few algorithms I used.  Some folks have been experimenting with "ideal"
> values for tcache_count and tcache_max too.
>
That isn't case, tcache counters slow things down and are unneccessary.
One could establish limits by using tricks in free.

I had different idea for representation now, so I will send it in
separate thread.

 
> > Finally for larger allocations I would use page-based logic as
> > mmaping/remapping/unmapping is about only way to actually decrease
> > memory footprint, I didn't try that much yet.
> 
> Note that mmap() itself is expensive (slow) and there may be a limit on
> how many discrete mapped regions a kernel[*] can support.
> 

For larger its more that one should be aware of paging as to handle
memory allocated from system. One such issue is that often big
allocation doesn't use pages at end. 
For returning memory to system I think that best strategy is something like
 hitting pages that weren't used last second by madvise(dont_need)