Myllyenko opened a new pull request, #4322:
URL: https://github.com/apache/logging-log4j2/pull/4322
`InstantPatternThreadLocalCachedFormatter` keyed its per-thread cache on a
`Long`: the epoch instant was extracted through a `Function<Instant, Long>` and
stored, boxed, into an `Object[]` held in a `ThreadLocal`. Both extractors,
`Instant::getEpochMillisecond` and `Instant::getEpochSecond`, return a
primitive, and epoch values are far outside the `Long` cache.
The extractor's box turns out to be scalar-replaced on a cache hit, since it
is unboxed straight into a local and never escapes. The box stored into the
`Object[]` does escape, so every cache miss allocated. A miss happens once per
distinct instant per thread, which for any thread logging at under a thousand
events per second is every single event.
Replace the extractor with a `ToLongFunction<Instant>` and the `Object[]`
with a small holder carrying a primitive `long` field next to the buffer.
`StringBuilderEncoder` documents a preference for keeping only JDK types in
thread locals, to avoid pinning a class loader in a web container, but that
does not apply here: this wrapper is only installed when
`Constants.ENABLE_THREADLOCALS` is set, and that flag is off for web
applications precisely to disable the thread locals which could leak.
Measured with the JMH benchmark added here, 1000 instants per operation:
before after
cacheMiss 24000.152 B/op, 153 GCs 0.141 B/op, no GC
cacheMiss 21744 +- 2404 ns/op 20274 +- 406 ns/op
cacheHit 0.038 B/op 0.040 B/op
cacheHit 5410 +- 70 ns/op 5696 +- 699 ns/op
Also close a cache-poisoning window while here: the cached epoch instant is
now cleared before the buffer is rewritten and restored only once formatting
succeeds, so a throwing formatter can no longer leave a later call reading a
half-written buffer. Note that merely moving the assignment after the
formatting does not fix this; it only changes which instant triggers it.
**INSERT HERE** a clear and concise description of what the pull request is
for along with a reference to the associated issue IDs, if they exist.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
