On Friday, 10 October 2014 at 08:45:38 UTC, Marc Schütz wrote:
Yes and quite notably so as GC.malloc can potentially trigger
collection. With concurrent GC collection is not a disaster
but it still affects the latency and should be avoided.
Is it just the potentially triggered collection, or is the
actual allocation+deallocation too expensive?
collection - for sure.
allocation+deallocation - maybe, I have never measured it. It is
surely slower than not allocating at all though.
Because the effects of the former can of course be reduced
greatly by tweaking the GC to not collect every time the heap
needs to grow, at the cost of slightly more memory consumption.
This is likely to work better but still will be slower than our
current approach because of tracking many small objects. Though
of course it is just speculation until RC stuff is implemented
for experiments.
Also let's note that extending existing chunks may result in
new allocations.
Yes. But as those chunks never get free'd it comes to O(1)
allocation count over process lifetime with most allocations
happening during program startup / warmup.
Hmm... but shouldn't this just as well apply to the temporary
allocations? After some warming up phase, the available space
on the heap should be large enough that all further temporary
allocations can be satisfied without growing the heap.
I am not speaking about O(1) internal heap increases but O(1)
GC.malloc calls
Typical pattern is to encapsulate "temporary" buffer with the
algorithm in a single class object and never release it, reusing
with new incoming requests (wiping the buffer data each time).
Such buffer quickly gets to the point where it is large enough to
contain all algorithm temporaries for a single request and never
touches GC from there.
In a well-written program which follows such pattern there are
close to zero temporaries and GC only manages more persistent
entities like cache elements.