stuart.marks at oracle.com
Mon Feb 3 22:01:27 UTC 2014
Nobody took the bait on this yet? :-)
Certainly there's a lot of semi-myth on this topic, on both sides. Here's my
source of mythology (or urban legend, as one might have it):
My concern here is not so much about leaking of the StringBuilder held in a
field, as Chris seemed to be responding to. I'd expect the ObjectInputStream to
be GC'd at some point along with any StringBuilders it contains references to.
I'm more concerned about defeating any optimizations that might occur if local
variables are converted to fields. The article referenced above mentions escape
analysis and possible stack allocation of locally-created objects. Offhand I
don't know if stack allocation occurs for any of the builders in
ObjectInputStream, but it certainly cannot occur if references are stored in fields.
It would good if there were some evidence we could discuss instead of myths and
urban legends. :-) Perhaps the original poster (Robert Stupp) can rerun some
benchmarks with and without the conversion from locals to fields. (Earlier in
the thread he posted some significant performance improvements, but my suspicion
is that most of that improvement came from the conversion to use Unsafe.)
I'm mindful that this may require a lot of effort. I think it would take a fair
bit of work to come up with a benchmark that shows any difference between the
two cases. I'm also mindful that one's intuition is often wrong.
On 1/31/14 5:53 PM, Vitaly Davidovich wrote:
> One would have to measure of course, but intuitively, it seems like a good
> change to reuse the stringbuilder. There's also the issue that the local
> stringbuilder before was default-sized, and for large content, it would
> generate even more garbage as the underlying array is expanded.
> The "temporary short lived allocations are cheap" is, unfortunately, a
> semi-myth, or at least somewhat misguided. It's true that individually they
> may be cheap, but they have macro effects. The higher the allocation rate,
> the more young GCs we get. Every young GC (on hotspot collectors, at least)
> is a STW pause. Bringing threads to a safepoint has some cost, especially if
> there are many of them on large many-core machines. With larger heaps these
> days, young gens tend to be larger as well. When GC runs, it trashes the cpu
> caches for the mutators, so when they resume, they may get cache misses. At
> each young GC, some objects may get promoted prematurely to tenured.
> So no, I wouldn't say it's quite inexpensive :). When you have no option but
> to allocate, it's nice to have collectors that can handle those necessary
> allocations well. Otherwise, if it's a perf sensitive area and avoiding
> allocations doesn't obfuscate or make the code significantly less
> maintainable, it's usually better to avoid allocations.
> Just my $.02
> Sent from my phone
> On Jan 31, 2014 2:06 PM, "Stuart Marks" <stuart.marks at oracle.com
> <mailto:stuart.marks at oracle.com>> wrote:
> On 1/31/14 10:46 AM, Chris Hegarty wrote:
> I think your patch can be split into two logical, independent, parts. The
> first is the use of unsafe to access the String UTF length. The seconds is
> to reuse, where possible, the existing StringBuilder instances,
> skipping of
> primitive/object reading/writing where applicable, and general cleanup.
> Since this is a very sensitive area I would like to consider these
> separately. To that end, I have taken the changes that are applicable
> to the
> latter, removed any subjective stylistic changes, and made some additional
> cleanup improvements.
> I agree with splitting the Unsafe usages so they can be reviewed
> separately. New
> and changed usage of Unsafe will require exacting scrutiny.
> In general, the cleanups and refactorings look fine.
> I have a question about the change to reuse StringBuilder instances. This
> replaces freshly allocated StringBuilders stored in local variables with
> reuse of a StringBuilder stored in a field of BlockDataInputStream, which
> in turn is stored in a field of ObjectInputStream. Thus, instead of
> creating of lots of temporaries that become gc-eligible very quickly, this
> creates a single long-lived object whose size is the high-water mark of
> the longest string that's been read. The change does reduce allocation
> pressure, but the point of generational GC is to make allocation and
> collection of temporaries quite inexpensive. Is this the right tradeoff?
More information about the core-libs-dev