Re: [Libunwind-devel] non-local return scalability bottleneck

[Milian Wolff]

On Sonntag, 6. Dezember 2015 11:31:17 CET Paul Pluzhnikov wrote:
> On Sun, Dec 6, 2015 at 11:03 AM, Arun Sharma <address@hidden> wrote:
> > Suspect the comment below is relevant. Perhaps try removing it and see
> > what happens?
> 
> I suspect that the dlclose resulting in unload, followed by dlopen of
> something else that lands in the same address range,  is what would
> break (by returning now stale info).

I've seen this bottleneck myself btw. in my heaptrack profiler - it costs both 
pure CPU performance as well as severely impacting the multi-thread 
performance.

I was thinking of writing a patch for an off-by-default cache with the 
capability to manually invalidate the cache whenever dlopen/dlcose gets called 
(which is rare, and detectable by a profiler). I never got around to write it 
though... If anyone could do that - it would be much appreciated!

That said, I did write this patch which is still not reviewed or accepted 
upstream:

https://lists.gnu.org/archive/html/libunwind-devel/2015-09/msg00005.html

@Luke: Could you try that patch and increase the cache size? That should 
already reduce the number of times dl_iterate_phdr gets called.

@Arun: Any chance you could review and land it?

Thanks
-- 
Milian Wolff
address@hidden
http://milianw.de

signature.asc
Description: This is a digitally signed message part.