LMMS Memory Manager

[diizy]

See post on mailing list for details.

Testing appreciated before merging.

[Sti2nd]

I'll just link to the mailthread for easy finding later http://linux-multimedia-studio-lmms.996328.n3.nabble.com/LMMS-Memory-Manager-td10076.html

[diizy]

66 files changed and counting... this is building up to a pretty massive pull request. I might have to write a whole brief just to document all the changes to the codebase done here...

Note about the current state of the branch (in case someone is testing this): the vol/pan knobs on instruments are temporarily broken... not a bug, just unfinished business. Will be addressed before this branch gets merged.

[tresf]

Holler if you want a windows build. I feel bad I haven't been helping with testing lately, but I can at least fire off a build for our testers... :)

[diizy]

Nah, not just yet, see: broken knobs. But yeah I'll let you know, thanks.

[diizy]

@tresf, the vol/pan functionality is restored now.

[tresf]

Ok, I tried to build this, but it appears it still needs some work for the terribly fussy win32 world:

I remember with the SWH ladspa C files there were quite a few Win32 forward-declaration workarounds but that is just speculation.

Linking CXX shared module amplifier.dll
CMakeFiles/amplifier.dir/objects.a(Amplifier.obj): In function `Effect::operator new(unsigned long long)':
/home/ubuntu/lmms/include/Effect.h:42: undefined reference to `MemoryManager::alloc(unsigned long long)'
CMakeFiles/amplifier.dir/objects.a(Amplifier.obj): In function `configManager::operator new(unsigned long long)':
/home/ubuntu/lmms/include/config_mgr.h:52: undefined reference to `MemoryManager::alloc(unsigned long long)'
/home/ubuntu/lmms/include/config_mgr.h:52: undefined reference to `MemoryManager::alloc(unsigned long long)'
/home/ubuntu/lmms/include/config_mgr.h:52: undefined reference to `MemoryManager::alloc(unsigned long long)'
CMakeFiles/amplifier.dir/objects.a(Amplifier.obj): In function `Effect::operator delete(void*)':
/home/ubuntu/lmms/include/Effect.h:42: undefined reference to `MemoryManager::free(void*)'
CMakeFiles/amplifier.dir/objects.a(Amplifier.obj): In function `AutomatableModel::operator delete(void*)':
/home/ubuntu/lmms/include/AutomatableModel.h:69: undefined reference to `MemoryManager::free(void*)'
collect2: error: ld returned 1 exit status
make[2]: *** [plugins/Amplifier/amplifier.dll] Error 1
make[1]: *** [plugins/Amplifier/CMakeFiles/amplifier.dir/all] Error 2
make: *** [all] Error 2

[diizy]

On 08/27/2014 11:31 PM, Tres Finocchiaro wrote:

Ok, I tried to build this, but it appears it still needs some work for
the terribly fussy win32 world:

Well, someone needs to step up to deal with this, or the next version of
LMMS will be Linux/Mac only...

[tresf]

I spent some time on it to no avail. I'm pretty sure the offending line is here (MM_OPERATORS):
https://github.com/diizy/lmms/blob/3bb2c50f77b0bed909206226fbfe6aee1531c8d3/include/AudioPort.h#L39

[diizy]

That's just a macro for overriding new/delete operators. It's necessary to make the class use the MemoryManager instead of doing direct heap allocations. It's used in many other classes as well.

[tresf]

It's used in many other classes as well.

Yes, I see that. This issue only occurs during the linking process, which isn't uncommon from what I've seen with MinGW.

Without much knowledge of the macro realm, could you explain a bit to me so that I can research some of the techniques used against the internet of things?

[diizy]

On 08/28/2014 02:15 AM, Tres Finocchiaro wrote:

It's used in many other classes as well.

Yes, I see that. This issue only occurs during the linking process,
which isn't uncommon from what I've seen with MinGW.

Without much knowledge of the macro realm, could you explain a bit to
me so that I can research some of the techniques used against the
internet if things?

A macro is a command for the C PreProcessor (CPP, not to be confused
with CPlusPlus).

What it does is basically just string replacement. When you #define a
macro, that means that anywhere in the code that macro is found, it is
replaced with the text in the #define, it won't be checked for any
syntax or program logic, it's just straight-up text replacement. Macros
are really more of a C thing and often kind of frowned upon by C++ purists.

Here, the macro simply adds operator overrides to the classes where it
is invoked. It overrides the new and delete operators. Usually these
operators call malloc(), which allocates memory on the heap, and this is
the thing we want to avoid, so the macro overrides them to call
MemoryManager::alloc() instead, which doesn't allocate memory on the
heap, but instead fetches memory from the pre-allocated memory pools.

[tresf]

Ok thanks. Some of the mingw posts I read mentioned the preprocessor and how it can cause linking issues. I have to read more, but that puts me in the right direction.

[tresf]

So the preprocessor stuff I'm stumbling on warns of something refereed to as "name mangling". MinGW has a page dedicated to this. Unfortunately the majority of it is above my head. Here's the article...

http://www.mingw.org/wiki/Interoperability_of_Libraries_Created_by_Different_Compiler_Brands

Edit: ... and another article... http://www.transmissionzero.co.uk/computing/building-dlls-with-mingw/

Seems we use this approach in export.h

[tresf]

@tobydox, should this be class EXPORT InstrumentPlayHandle ? It's the only way I could get compilation for Win32/64 to succeed but I'm just taking guesses here.

[tresf]

For anyone willing to test (@musikBear, @Sti2nd, @StakeoutPunch) here are windows builds of master with new MemoryManager.

https://github.com/tresf/lmms/releases/tag/v1.0.990

[Sti2nd]

Played around in 64 bit, no bugs. A quick question, is the CPU meter fixed? My processor is at 22% at maxiumum,

and LMMS shows the CPU

almost full on unfa-Spoken. So it is clearly wrong...?

[tresf]

and LMMS shows the CPU almost full on unfa-Spoken. So it is clearly wrong...?

Looks fine over here. :)

[diizy]

On 08/28/2014 03:37 PM, Stian Jørgensrud wrote:

Played around in 64 bit, no bugs. A quick question, is the CPU meter
fixed? My processor is at 22% at maxiumum,
image
https://cloud.githubusercontent.com/assets/6373430/4075463/08765b02-2eb0-11e4-80cf-c91f83731626.png
and LMMS shows the CPU
image
https://cloud.githubusercontent.com/assets/6373430/4075453/f2dd8fb8-2eaf-11e4-920b-b1b9b113d96d.png
almost full on unfa-Spoken. So it is clearly wrong...?

Nope, it's not wrong. The LMMS CPU meter does not necessarily correlate
with CPU usage stats. LMMS shows the amount of CPU used from maximum
amount of CPU resources available to LMMS, based on how much time it
takes to render the period. For example, if your period size is set to
5ms, and rendering the period takes 2.5ms, then that translates to 50%
CPU usage.

[Sti2nd]

Oooh

[exTunes]

I posted a bug here: #1108

[diizy]

Ok, did a sync and rebase, to get Toby's GUI-less rendering fix in.

So now it should be possible to benchmark this branch against the current master. I'd be very interested to know if there are any quantifiable performance benefits.

[tresf]

So this isn't exactly a benchmark, but I compared memory usage on Win7 x64:

Memory Usage - unfa-Spoken.mmpz

• 1.0.3 - 573MB
• 1.0.991 - 697MB

There're many reasons why this shouldn't be considered even a reasonably accurate benchmark: First - It compares an EOL branch to a Bleeding edge branch so there are way too many changes to make any argument that the MemoryManager changes have contributed to this. Second - Memory efficiencies shouldn't be measured by the memory overhead, but rather the engine efficiency as a whole (including spikes, io tomes, etc). However, despite all of this I still wanted to share some early findings. 🐰

[exTunes]

With the Memory Manager Branch, I can master a rendered track in a new file.
Usually LMMS crashes (with more than 7 GB RAM consumed) after playing back a sample track for one minute.. even without any effects!

[diizy]

On 08/30/2014 04:29 PM, Tres Finocchiaro wrote:

So this isn't exactly a benchmark, but I compared memory usage on Win7
x64:

Memory Usage - unfa-Spoken.mmpz

• |1.0.3| - 573MB
• |1.0.991| - 697MB

There's many reasons why this shouldn't be considered even a
reasonably accurate benchmark: First - It compares an EOL branch to a
Bleeding edge branch so there are way too many changes to make any
argument that the MemoryManager changes have contributed to this.
Second - Memory efficiencies shouldn't be measured by the memory
overhead, but rather the engine efficiency as a whole (including
spikes, io tomes, etc). However, despite all of this I swill wanted to
share.

Oh no, it's probably fairly accurate.

Memory usage /is /going to increase quite a bit with the MemoryManager -
this is because it allocates a huge chunk of memory at startup, which is
then used throughout the program... Also, the memory is then used as
"chunks" of 64 bytes, so every object's memory footprint gets rounded up
to the nearest 64 bytes, meaning that if you use the MM to allocate
memory for an object that is 65 bytes in size, it will allocate 128
bytes for it.

All of this causes some waste in memory, but here's the thing... I don't
really care about that. RAM is cheap these days, tradeoff of higher
memory usage -> better performance is one that I'm more than willing to
make.

[eagles051387]

What you mentioned about ram being cheap. I wonder if someone has an SSD a
chunk of memory so to speak can be allocated on the SSD

On Sun, Aug 31, 2014 at 8:11 AM, Vesa V notifications@github.com wrote:

On 08/30/2014 04:29 PM, Tres Finocchiaro wrote:

So this isn't exactly a benchmark, but I compared memory usage on Win7
x64:

Memory Usage - unfa-Spoken.mmpz

• |1.0.3| - 573MB
• |1.0.991| - 697MB

There's many reasons why this shouldn't be considered even a
reasonably accurate benchmark: First - It compares an EOL branch to a
Bleeding edge branch so there are way too many changes to make any
argument that the MemoryManager changes have contributed to this.
Second - Memory efficiencies shouldn't be measured by the memory
overhead, but rather the engine efficiency as a whole (including
spikes, io tomes, etc). However, despite all of this I swill wanted to
share.

Oh no, it's probably fairly accurate.

Memory usage /is /going to increase quite a bit with the MemoryManager -
this is because it allocates a huge chunk of memory at startup, which is
then used throughout the program... Also, the memory is then used as
"chunks" of 64 bytes, so every object's memory footprint gets rounded up
to the nearest 64 bytes, meaning that if you use the MM to allocate
memory for an object that is 65 bytes in size, it will allocate 128
bytes for it.

All of this causes some waste in memory, but here's the thing... I don't
really care about that. RAM is cheap these days, tradeoff of higher
memory usage -> better performance is one that I'm more than willing to
make.

—
Reply to this email directly or view it on GitHub
#1088 (comment).

Jonathan Aquilina

[diizy]

On 08/31/2014 09:24 AM, eagles051387 wrote:

What you mentioned about ram being cheap. I wonder if someone has an
SSD a
chunk of memory so to speak can be allocated on the SSD

That's for the OS to handle with swap partitions and such. We're not
going to do any swapping in the LMMS codebase, I can't see it being
beneficial in any way... we need quick memory access, and swapping seems
to me like something that would be horrible for RT-safety.

[diizy]

So, I think I'm just going to merge this branch now. There's so many commits here, getting troublesome to keep this in sync with master, etc... so any performance issues can be profiled and fixed later - if we bother to work much more on 1.2 anymore, since it seems we could just skip right into 2.0...

[jasp00]

This memory manager should be dropped. I do not think it is better than the allocator in glibc. MemoryManager uses locks, while glibc's implementation is "without lock contention", according to the source. MemoryManager uses a slow algorithm to find free chunks.

The use of a custom manager should be centered on objects created by the mixer, and leave out other objects such as SampleBuffer. Either a more developed general purpose allocator should be used, TLSF could be an option, or a manager specific to LMMS' needs should be designed.

[midi-pascal]

👍 for TLSF that I use in the top 1 commercial video games sound engine.

[jasp00]

I have been looking at TLSF and glibc malloc. They both use locks for multiple threads. Without locks, an inefficient allocation method is needed. One conclusion is that LocklessList is not lockless, because new and delete are on different threads.

Allocation methods that do not consider a maximum usage incur a memory pool extension, which affects real-time processing. General purpose allocators face fragmentation, which makes the pool extension more likely.

So the way to be lockless is to use structures with a maximum number of same-sized elements, something like m_items in MixerWorkerThread. Using bit arrays to detect free elements may help a little with efficiency.