Issue #16648 has been reported by bpowers (Bobby Powers).

----------------------------------------
Feature #16648: improve GC performance by 5% with builtin_prefetch
https://bugs.ruby-lang.org/issues/16648

* Author: bpowers (Bobby Powers)
* Status: Open
* Priority: Normal
----------------------------------------
The mark phase of non-incremental major GC is (I believe) dominated by pointer chasing.  One way we can improve that is by prefetching cachelines from memory before they are accessed, to reduce stalls.  I did some experimenting, and the following patch reduces the time spent on a full GC from ~ 950 milliseconds to ~ 900 milliseconds, a small but stable improvement.  I would love if additional folks have other benchmarks (or could point me at them) to see if this holds up beyond the web service I tested, and whether something like this could be considered for inclusion.

I also attempted a more "principled" approach based on an optimization described in the GC handbook: putting a FIFO queue in front of the mark stack, and prefetching addresses as they enter the queue.  However, I wasn't able to see any performance improvement there despite testing a number of queue sizes from 4 to 64.  Its possible I implemented this wrong, or misjudged the access patterns (if e.g. the memory of a VALUE is accessed before push_mark_stack is called, it would invalidate this approach).  The code for that alternative is here: https://github.com/bpowers/ruby/commit/d790d0c15047c36c23850a112093fe0e32fd3262

---Files--------------------------------
0001-gc-prefech-objects-seems-to-improve-full-GC-performa.patch (2.29 KB)


-- 
https://bugs.ruby-lang.org/

Unsubscribe: <mailto:ruby-core-request / ruby-lang.org?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>