Feature #10137: Introducing Incremental GC algorithm

• Status: Closed
• Assignee: ko1 (Koichi Sasada)
• Target version: 2.2.0

[ko1 (Koichi Sasada)]

Abstract

Introduce incremental GC algorithm to reduce pause time of major/full
GC. This ticket includes design and implementation note and a working
patch.

Background and problem

Ruby 2.1 uses generational GC algorithm (named RGenGC) to improve GC
throughput. Genrational GC algorithm separates existing objects into
young generation objects and old generation objects. At the most of GC
timing, GC only marks young generation objects (called minor GC). If
there are no enough memory, then mark all of objects (called major GC or
full GC). Minor GC is dramatically fast than major GC. So that total
throughput of application improves (10% improvement in my RDoc benchmark,
[ruby-list:49896] reported that GC intensive application is 6 times
faster!).

However, major GC is needed periodically and it pauses same time as GC
on Ruby 2.0 and before. This problem hits response time intensive
application such as web application.

Proposal

Introduce Incremental GC algorithm for major GC.

Incremental GC algorithm is well-known GC algorithm to reduce GC pause
time. Scatter GC (marking) process in several phases and run processes
interleaving with Ruby's execution process. This is similar to current
lazy sweep algorithm (in fact, lazy sweep is a half part of incremental
GC algorithm).

Running ruby process with marking phase, it is possible to introduce
critical bug because marked objects can points un-marked objects (on the
incremental GC terminology, marked objects are called "Black" objects
and un-marked objects are called "White" objects). Such un-marked
objects can be left in un-marking and be swept.

# if `ary' is already marked, and obj is not marked
ary[0] = obj
obj = nil # no-body except `ary' refers `obj'

To prevent such destructive bug, we can use write barriers to detect
such "marked objects" to "un-marked objects". We can care about such
case.

Yes, MRI/CRuby has "WB-unprotected" objects such objects does not have
write barriers because of compatibility or implementation issues. To
care about such WB-unprotected objects, we need to traverse all of
living WB-unprotected objects at a time in the last of incremental
marking. This is new extending idea against traditional incremental GC
algorithm (at least I surveyed).

Deisgn and implementation details are here:
http://www.atdot.net/fp_store/f.p61can/file.data-incremental-gc.pdf

Maybe a diagram at page 10 will help you to understand the flow of all
GC process.

Code is here:
https://github.com/ko1/ruby/tree/rincgc

Compare with trunk:
https://github.com/ko1/ruby/compare/rincgc

Implementation note

WB-unprotected bitmap

As I said, we need to check all of living WB-unprotected objects at the
last of incremental marking phase. To do it lightweight, introduce
WB-unprotected bitmap intead of specific bit in RBasic::flags.

We can get all living WB-unprotected objects with the following pseudo
code:

bits = mark_bits[i] & wb_unprotected_bits[i];
while (bits) {
  if (bits & 1) do something.
  bits >>= 1;
}

4 age promotion

Because we don't need to use WB-protected bit in RBasic::flags, we have
another 1 bit in RBasic::flags. To utilize this bit, we add age of an
object with exsiting promoted bit. Rename FL_WB_PROTECTED to
FL_PROMOTED0 and FL_PROMOTED to FL_PROMOTED1.

These two bits represent object's age (0 to 3) and 3 means OLD objects.

Write barriers

We extend write barriers to detect such [marked objects -> un-marked
objects] reference in incremental GC. It introduces some overhead.

Evaluation

Benchmark results on real Linux machine (*1)
http://www.atdot.net/sp/raw/yekban
*1: Intel(R) Xeon(R) CPU E5335 @ 2.00GHz, 4GB memory, 2.6.32-5-amd64 (Debian squeeze)

In most of case, there are only a few (~5%) performance down.
Incremental GC introduces some overhead. But I think they are
acceptable speed-down.

Discouse benchmark (only on Virtualbox VM, so accuracy is not good)
http://www.atdot.net/sp/raw/g9uban

We can recognize reducing worst case seconds.

TODO

(1) Clean up codes

Now, we can not disable incremental GC codes and generational GC codes.
We need to add ability to enable/disable features with macros.

(2) Tuning parameters

Now the parameters are fixed in codes. mruby already have tuning
parameters for incremental GC (matz said they are from Lua),
"GC.interval_ratio" and "GC.step_ratio". We can import these functions
(or making another interface to tell).

(3) Enter GC/ Exit GC internal events

This patch also includes function "gc_enter()" and "gc_exit()" which set
and reset a "doing GC" flag.

If we introduce internal event to hook these functions, we can measure
exact GC pause time (and mutators time).

Summary

This feature proposal is to introduce incremental GC algorithm with working code.
Incremental GC algorithm reduce application's pause time of major GC.

Any feedback are welcome!

Thanks,
Koichi

[ko1 (Koichi Sasada)]

• Description updated (diff)

Add an abstract section.

[mrkn (Kenta Murata)]

• Description updated (diff)

[hsbt (Hiroshi SHIBATA)]

I evaluated running time for test-all.

ref. https://twitter.com/k_tsj/status/500523453028388864

rincgc branch seems 2 times slower than trunk when sequential running.
I tested 3 times. it's same results.

trunk

ruby -v: ruby 2.2.0dev (2014-08-17 trunk 47206) [x86_64-darwin14]
make test-all TESTS='-j4' 291.72s user 65.89s system 241% cpu 2:27.94 total

ruby -v: ruby 2.2.0dev (2014-08-17 trunk 47206) [x86_64-darwin14]
make test-all 336.31s user 62.33s system 83% cpu 7:55.63 total

rincgc branch

ruby -v: ruby 2.2.0dev (2014-08-15 trunk 47192) [x86_64-darwin14]
make test-all TESTS='-j4' 359.57s user 74.27s system 213% cpu 3:23.33 total

ruby -v: ruby 2.2.0dev (2014-08-15 trunk 47192) [x86_64-darwin14]
make test-all 878.35s user 65.17s system 90% cpu 17:22.43 total

[ko1 (Koichi Sasada)]

I had accidentally added GC.verify_internal_consistency method for each
test case (in after_tear_down.

After remove it,

trunk : 798.756056s
before removal: 1981.145346s
after removal : 845.399831s

Not so bad.

[ko1 (Koichi Sasada)]

FYI: here is a top slow tests with data.
http://www.atdot.net/fp_store/f.zdsfan/file.copipa-temp-image.png

Typically, GC.start slows down.

[matz (Yukihiro Matsumoto)]

Go ahead. We need to experiment in real use cases.

Matz.

[ko1 (Koichi Sasada)]

• Status changed from Open to Closed
• % Done changed from 0 to 100

Applied in changeset r47444.

---

• gc.c: add incremental GC algorithm. [Feature #10137]
  Please refer this ticket for details.
  This change also introduces the following changes.
  • Remove RGENGC_AGE2_PROMOTION and introduce object age (0 to 3).
    Age can be count with FL_PROMOTE0 and FL_PROMOTE1 flags in
    RBasic::flags (2 bit). Age == 3 objects become old objects.
  • WB_PROTECTED flag in RBasic to WB_UNPROTECTED bitmap.
  • LONG_LIVED bitmap to represent living objects while minor GCs
    It specifies (1) Old objects and (2) remembered shady objects.
  • Introduce rb_objspace_t::marked_objects which counts marked
    objects in current marking phase. marking count is needed to
    introduce incremental marking.
  • rename mark related function and sweep related function to
    gc_(marks|sweep)_(start|finish|step|rest|continue).
  • rename rgengc_report() to gc_report().
  • Add obj_info() function to get cstr of object details.
  • Add MEASURE_LINE() macro to measure execution time of specific line.
  • and many small fixes.
• include/ruby/ruby.h: add flag USE_RINCGC.
  Now USE_RINCGC can be set only with USE_RGENGC.
• include/ruby/ruby.h: introduce FL_PROMOTED0 and add FL_PROMOTED1
  to count object age.
• include/ruby/ruby.h: rewrite write barriers for incremental marking.
• debug.c: catch up flag name changes.
• internal.h: add rb_gc_writebarrier_remember() instead of
  rb_gc_writebarrier_remember_promoted().
• array.c (ary_memcpy0): use rb_gc_writebarrier_remember().
• array.c (rb_ary_modify): ditto.
• hash.c (rb_hash_keys): ditto.
• hash.c (rb_hash_values): ditto.
• object.c (init_copy): use rb_copy_wb_protected_attribute() because
  FL_WB_PROTECTED is moved from RBasic::flags.
• test/objspace/test_objspace.rb: catch up ObjectSpace.dump() changes.