Feature #20351
closedOptionally extract common GC routines into a DSO
Description
UPDATE: Based on feedback on the original PR (thank you @katei (Yuta Saito) and @nobu (Nobuyoshi Nakada)) we have
changed our approach to this project.
An updated PR can be found here: [GH #10456]
NOTE: This proposal does not change the default build of Ruby, and therefore
should NOT cause performance degradation for Ruby built in the usual way
Our long term goal is to standardise Ruby's GC interface, allowing alternative
GC implementations to be used. This will be acheived by optionally building
Ruby's GC as a shared object; enabling it to be replaced at runtime using using
LD_LIBRARY_PATH
. eg:
LD_LIBRARY_PATH=/custom_gc_location ruby script.rb
This ticket proposes the first step towards this goal. A new experimental build
option, --enable-shared-gc
, that will compile and link a module into the built
ruby
binary as a shared object - miniruby
will remain statically linked to
the existing GC in all cases.
Similar methods of replacing functionality relied on by Ruby have
precedent. jemalloc
uses LD_PRELOAD
to replace glibc
provided malloc
and
free
at runtime. Although this project will be the first time a technique such
as this has been used to replace core Ruby functionality.
This flag will be marked as experimental & disabled by default.
The PR linked from this ticket implements the new build flag, along with the
absolute minimum code required to test it's implementation (a single debug
function).
The implementation of the new build flag is based on the existing implementation
of --enable-shared
and behaves as follows:
-
--enable-shared --enable-shared-gc
This will build both
libruby
andlibrubygc
as shared objects.ruby
will
link dynamically to bothlibruby
andlibrubygc
. -
--disable-shared --enable-shared-gc
This will build
librubygc
as a shared object, and buildlibruby
as a
static object.libruby
will link dynamically tolibrubygc
andruby
will
be statically linked tolibruby
. -
--disable-shared-gc
This will be the default, and when this case is true the build behaviour
will be exactly the same as it is currently. ie. the existing Ruby GC will be
built and linked statically into eitherruby
orlibruby.so
depending on
the state of--enable-shared
.
We are aware that there will be a small performance penalty from moving the GC
logic into a shared object, but this is an opt-in configuration turned on at
build time intended to be used by experienced users.
Still, we anticipate that, even with this configuration turned on, this penalty
will be negligible compared the the benefit that being able to use high
performance GC algorithms will provide.
This performance penalty is also the reason that this feature will be disabled
by default. There will be no performance impact for anyone compiling Ruby in
the usual manner, without explicitly enabling this feature.
We have discussed this proposal with @matz (Yukihiro Matsumoto) who has approved our work on this
project - having a clear abstraction between the VM and the GC will enable us to
iterate faster on improvements to Ruby's existing GC.
Motivation¶
In the long term we want to provide the ability to override the current Ruby GC
implementation in order to:
- Experiment with modern high-performance GC implementations, such as Immix, G1,
LXR etc. - Easily split-test changes to the GC, or the GC tuning, in production without
having to rebuild Ruby - Easily use debug builds of the GC to help identify production problems and
bottlenecks without having to rebuild Ruby - Encourage the academic memory management research community to consider Ruby
for their research (the current work on MMTk & Ruby is a good example of
this).
Future work¶
The initial implementation of the shared GC module in this PR is deliberately
small, and exists only for testing the build system integration.
The next steps are to identify boundaries between the GC and the VM and begin to
extract common functionality into this GC wrapper module to serve as the
foundation of our GC interface.
Who's working on this¶
Updated by eightbitraptor (Matt V-H) 8 months ago
Based on feedback from @katei (Yuta Saito) and @nobu (Nobuyoshi Nakada) we have changed our approach to this project.
An updated PR can be found here: [GH #10456]
Instead of building the Ruby GC as a shared object in order to override it using LD_LIBRARY_PATH
/LD_PRELOAD
et al, and after experimenting with some alternative approaches we have decided to use an approach based around the dlopen
wrappers provided in dln.c
. This provides several benefits that our existing implementation does not.
- It does not require integration of another shared object into the build system. As pointed out by @katei (Yuta Saito), this is a complex and error-prone task, with added complexity implications for committers building and testing new features on Ruby, as well as an explosion of new CI tasks required to test all variants.
- It doesn't expose the GC as a shared object, with the potentially misleading consequence of users assuming they can link against the Ruby GC independently; as they currently can with
libruby
- It runs on more platforms than our initial approach which was targeted specifically at Linux and MacOS. The use of existing functionality in
dln.c
should allow development of this feature on any platform that supports C extensions.
The approach taken here is as follows:
- Enable this feature by configuring with
--with-shared-gc
- When enabled, Ruby will check for the presence of the environment variable
RUBY_GC_LIBRARY_PATH
- If that path exists and points to a valid shared object
- Open the shared object (using
dln_open
), and map theInit_GC
function to a function map - call
Init_GC
as part ofrb_objspace_alloc
when initialising the GC
- Open the shared object (using
- If that path does not point to a valid shared object
- Degrade gracefully to initialising the existing Ruby objspace and GC
When Ruby is configured without the --with-shared-gc
flag, no extra code is compiled and so existing behaviour is maintained without any performance penalty.
Enabling this feature, but not populating RUBY_GC_LIBRARY_PATH
will incur a small performance penalty due to the overhead of using getenv
to check for the environment variable.
The use of dln.c
rather than using dlopen
directly, should enable this feature to be supported on all platforms that support loading shared libraries, rather than the initial implementation, which explicitly only supported Linux and MacOS.
Updated by eightbitraptor (Matt V-H) 7 months ago
Closed by https://github.com/ruby/ruby/pull/10456
Updated by eightbitraptor (Matt V-H) 7 months ago
- Status changed from Open to Closed
Updated by ko1 (Koichi Sasada) 7 months ago
Please note that we are working on ractor local GC and I'm not sure we can fix "GC interface".
Updated by eightbitraptor (Matt V-H) 7 months ago
Do you have an idea of a timeline for Ractor local GC? Is there an open feature request I can read (I've searched on Redmine and I can't find one).
It's hard for us to assess any collisions between this work and Ractor local GC at the current time as I'm unsure about the implementation details.
Our intention is to allow the GC to be overloaded in such a way that multiple heaps can be initialised and managed independently, in isolation from each other or the VM, so potentially each Ractor could use this feature to manage it's own heap.
To be clear this feature request and the associated PR are only providing some of the base mechanism to allow us to experiment. We have not committed to an approach at this stage.
With our current knowledge I believe it is possible to accommodate per-ractor GC and we will be careful to think about this in our design. In order to do this it would be very useful to have some guidance on the current per-ractor GC work.