"[email protected]" <[email protected]> writes:
> What happens if your run it a third time at the end? (The question
> is related to the fact that there appears to be transition states
> between monomorphic and megamorphic call sites, which might lead to
> an explanation.)
Same results, but your comment jogged my reading and analysis is what I
believe is the right direction.
This is a potentially grandiose claim, but I believe the inverse speedup
is due to contention caused by JVM type profiling in megamorphic call
sites. Unfortunately -XX:-UseTypeProfile doesn’t appear to turn off
type profiling itself, so I can’t prove this claim definitively without
going even further down the rabbit hole and grubbing through the JVM
source code.
To back up this claim, I present the following modified `conj*`:
(defn conj*
[coll x]
(let [a (long-array 32)]
(dotimes [n 5] (dotimes [i 32] (aset a i n)))
(clojure.lang.RT/conj coll x)))
And the resulting profiling numbers:
list-conj* : map-ms: 42.0, pmap-ms 24.8, speedup 1.69
cons-conj* : map-ms: 39.7, pmap-ms 25.1, speedup 1.58
Adding busy-work (and an extra allocation!) to the loop improved the
speed-up, I believe by decreasing the relative portion of the call
execution time during which the callsite type profile information is
being updated. If I’m correct, the `Cons` and `PersistentList` `.cons`
implementations are so tight that in their base versions the type
profiling forms a significant enough portion of the total call that the
updates are in frequent conflict. Adding busy-work to the `conj*` call
adds some jitter which prevents them from contending quite as
frequently.
I’m not sure what the next steps are. Open a bug on the JVM? This is
something one can attempt to circumvent on a case-by-case basis, but
IHMO has significant negative implications for Clojure’s concurrency
story.
-Marshall
--
You received this message because you are subscribed to the Google
Groups "Clojure" group.
To post to this group, send email to [email protected]
Note that posts from new members are moderated - please be patient with your
first post.
To unsubscribe from this group, send email to
[email protected]
For more options, visit this group at
http://groups.google.com/group/clojure?hl=en
