On 04/29/2014 03:43 PM, Ricardo Nabinger Sanchez wrote:
On Mon, 28 Apr 2014 15:08:16 +0400
Antony Dovgal <[email protected]> wrote:
I suspect that this is caused by memory fragmentation issues, though I've no
idea how to prove it (or at least all my ideas involve complex stats and are
quite inefficient).
You could dump your pointers and their sizes, and plot that information
graphically, including relevant info from jemalloc (map bases and their
sizes as well). If it's heavily fragmented, it will be apparent.
Dumping millions of pointers and hashtables (I use Judy arrays a lot) to plot
the data - that's what I meant by complex stats, yes =).
If that is indeed your case, and you can modify your data container,
perhaps switching to index-based access could solve your problem. That
would require you to forget about pointers and instead use an index, so
that your data container is free to move data around and shrink the
memory in use. There is a lot of management and memory copying,
meaning the impact of such a change should be quite perceptible.
Of course, this is no trivial change. Dealing with long- and
short-lived objects is challenging. :-)
Yeah, it's quite non-trivial.
Thanks for the suggestion, though, I'll keep it in mind.
--
Wbr,
Antony Dovgal
_______________________________________________
jemalloc-discuss mailing list
[email protected]
http://www.canonware.com/mailman/listinfo/jemalloc-discuss
