[Vm-dev] Garbage Collection (was Re: [Pharo-dev] Discussing the roadmap)
btc at openinworld.com
Sun Dec 3 05:33:14 UTC 2017
Hi Eliot, Clement,
On 7 July 2017 at 00:41, Eliot Miranda <eliot.miranda at gmail.com> wrote:
> > - Better support for large heaps (GC tuning API, incremental GC).
> > Pharo 64 bit is now able to manage large heap. However better
> > performance can be proposed by offering better settings for the
> > different GC zone.
> The most important thing here is the incremental GC. Spur has a generation
> scavenger that collects garbage in newly created objects (new space),
> and a mark-compact collector that collects and compacts garbage in old space.
> Right now on my 2.3GHz MacMini doing normal development, the generation
> scavenger causes pauses of 1ms or less, and the mark-compact collector
> than causes pauses of around 200ms. Both account for about 0.75% of
> entire execution time (1.5% total), so the scavenger is invoked frequently
> and the pauses that it creates are not noticeable. But the occasional
> pauses created by the mark-compact collector /are/ noticeable,
> especially in games and music.
> The idea for the incremental collector is to implement a mark-sweep or
> mark-sweep-compact collector for old space that works incrementally,
> doing a little bit of work on each invocation, probably immediately after a scavenge.
> It is intended to avoid the long pauses caused by the non-incremental
> mark-compact collector and make the system more suitable for games, music, etc.
"An alternative implementation, oft-used in Lisp systems, is to add a
read barrier to all object access, and mark objects as forwarders.
This can be used to implement a >>>lazy copying garbage collection<<<<
where objects are copied from one semi-space to another in parallel to the
main program (the “mutator”). To become, or move an object one replaces the
object’s header or first field with a forwarding pointer to the desired target
or copy in a new location, marking the “corpse” as forwarded. The program
checks the forwarded flag on each access. If there is hardware support,
as in a Lisp machine, this can work well. But without hardware support,
and like the object table representation, it has costs, slowing down
program execution due to the scattering of forwarding checks and
forwarding pointer follows throughout program execution."
I'm curious... Given we now have forwarders with Spur, are we
already sufficiently paying the cost of forwarding checks that a lazy copying
garbage collector might be a feasible form of incremental garbage collection?
I presume "parallel to the main program" means garbage collection occuring
in a separate thread to the main vm thread, potentially resulting in
very low main program pause times for garbage collection.
I found this a useful summary of the terminology...
and I'm curious how our planned Incremental CG fits those categories.
That article got me contemplating our performance constraint of the VM only
operating in only in a single native thread. Even though GC is a only
a few percent of performance, I wondered what a concurrent GC might
look like for us.
I found this video describing concurrent GC in Go (ignore first 11:20
and the second half was not so interesting)
where they present some interesting charts of their concern with latency pauses.
(btw they reference a multi-language GC latency benchmark
And for balance of that I found...
And then my mind wandered around implementation details of concurrent
To organise and quiet my thoughts I needed to put pen to paper, so I
thought sharing that
might stir thoughts for others. Probably naive and please excuse the
brain dump format...
Considering two threads...
* Main program thread "MP"
* Garbage collection thread "GC"
with object-space shared between them, consisting of objects split in
object-header & object-body...
1. Only "MP" mutates the object-body, updating slots creating new edges
in the object-graph, and relocating objects in memory using forwarders.
This rule avoids potential race conditions without needing to add
synchronisation code affecting performance of "MP".
2. Concurrently "GC" performs a Marking Phase by following the object-graph
tricolour tagging objects gray & black. it needs to mutate the object-header,
in a synchronised way something like...
a. Load object-header from shared object-space into local variable
H <== object-header
b. Modify header into another local variable
H' <== H + updated GC color bits
c. Atomic compare-and-swap H' back into object-space
object-header <== if H then H'
"MP" gets priority. Conflicts presumed rare.
Then considering object mutation by "MP" concurrent/overlapped with
3. TLDR; see 4.
In old-space if 'mutated' object is linked to a 'target' object
a. if 'mutated' isBlack, "MP" marks 'target' gray (to be later
processed by "GC").
b. if 'mutated' isGray, overlapped 'mutated' gray->black by "GC"
remains white would break tricolor invariant (is it a credible
i. mark 'target' gray, the state anyway in case 'mutated'
gray -> black
ii. re-mark 'mutated' gray, i.e. normally gray->gray, and
iii. optimisticly just check after mutation if color changed
cheaper than writes and conflicts presumed rare) then
set color to gray.
c. if 'mutated' isWhite, options...
i. do nothing and let "GC" reach 'target' normally, as long
as overlapped white->black cannot occur
ii. if white->black possible, do same as (3.b) "MP" marks
'target' gray, to be later processed by "GC".
4. Overall simplification of (3.) might be...
"MP" checks for any color change during mutation, and only then
marks 'mutated' object gray.
How expensive would such a check be? Presumed marking is infrequent and
can be done safely like (2.)
5. Eden/survivor space is ignored by "GC" thread. No point in adding
work to the gray set
until survivors filtered. Stop the world scavenging done as normal by "MP",
only marking objects gray when they are moved to old-space.
Post scavenging options...
a. Resume the world and leave it for "GC" to process these
recently grayed objects.
b. Keep world stopped for "MP" to complete marking, emptying gray set to
transition to Sweep Phase, at which point "MP" resumes the world.
It doesn't matter if subsequently objects are added to the gray set,
since existing white objects can never again be referenced by "MP".
6. Sweeping can be safely done by "GC" since white-objects are
unreachable from "MP".
"GC" can also take time to determine an optimum page P to compact and
then (per 1.) passes to "MP" via a relocation-queue the objects to be
relocated using forwarding pointers. "GC" could even spend extra time
to determine the optimum relocation-destination without impacting the
performance of "MP". When "MP" empties the relocation-queue, "GC" starts
on the next Marking Phase.
7. Now a question remains about multi-threaded flattening of
If two threads simultaneously perform an identical transform from...
someObject-slot --> forwarder-b --> finalObject-c
someObject-slot --> finalObject-c
does it matter that these operation may be done twice overlapping?
a. One mitigation could be for "GC" to identify forwarders to be flattened
and queue them for "MP" to process (reuse the compaction-queue). This
is work that "MP" would need to do anyway, but brings it forward to be
dealt with at a convenient time.
b. I guess "MP" and "GC" could play nice together if when encountering
a slot containing with a forwarding pointer they both do
something similar to (2.) like...
i. Load object-slot from shared object-space into local
variable S <== slot
ii. Set local variable S' <== flattened/followed pointer
iii. Atomic compare-and-swap S' back into object-space,
(object-slot <== if S then S')
- "MP" gets priority but presume conflicts rare anyway.
- While this violates rule (1.) the presumed frequency of
encountering forwarding pointers is low for "MP", so performance
should not be affected. Indeed by "GC" pre-emptively
the frequency "MP" sees reduces.
8. The release of the compacted page back to the OS is held up by
Forwarders are part of the graph followed in the Marking Phase they get
marked gray/black just like objects if they are referenced. After
fully flattened they are skipped by Marking and end up marked white
just like any other object. Once all forwarders are released, the
page is released
to the OS. If "GC" can effectively flatten pointers concurrently
with "MP" during its
normal Marking Phase, then pages would be released back to the OS
in a timely manner.
Now one thing I am curious about how the GC tri-color marking is implemented.
it describes how Spur's object header has three bits for GC.
* isMarked (I presume this means marked "black")
Do the bits imply the gray set is not stored in a separate data
structure on the heap, but rather distributed in-place, which I guess
would require multiple passes through memory to grow the gray set?
So this is not an area I'm set up to seriously work on, but I remain curious
and hopefully its a useful seed discussion for others.
P.S. At https://clementbera.wordpress.com/2017/09/19/vm-learning-memory-management/
nice to hear that you have Sophie (I presume) continuing with the VM.
More information about the Vm-dev