Review Request: UseNUMAInterleaving

[Deneau, Tom]

I always assumed that any application that spun up a small enough
number of threads that all the threads could execute on one node
would get hurt if any part of the heap was interleaved.

-- Tom

> 
> Do we have a list of such applications?  If not, we could add an
> analysis to the performance
> team's roadmap.
> 
> Paul
> 
> On 5/17/11 12:58 PM, Igor Veresov wrote:
> > UseNUMA already has this mechanism. If you turn it on it will
> > deactivate itself if the machine is not NUMA. The problem with turning
> > it on by default is that it can be not necessarily beneficial for all
> > applications (however for most of them it is).
> >
> > igor
> >
> > On 5/17/11 5:04 AM, Paul Hohensee wrote:
> >> I believe you can get memory geometry info from Solaris on both sparc
> or
> >> intel.
> >> Don't know about Windows or Linux. Anyone??
> >>
> >> Paul
> >>
> >> On 5/17/11 5:12 AM, Andrew Haley wrote:
> >>> On 17/05/11 00:18, Paul Hohensee wrote:
> >>>
> >>>> I suggest you just use -XX:+UseNUMA rather than adding a new
> >>>> flag. UseNUMA seems generic enough to cover whatever implementation
> >>>> is best on a particular platform.
> >>> Indeed. I am worried by the proliferation of command-line
> >>> optimization flags that have obscure meanings that are difficult to
> >>> understand by anyone unfamiliar with a particular computer
> >>> architecture. This sort of tweakage might be mostly used by people
> >>> doing benchmarking.
> >>>
> >>> When should this flag *not* be used? To make it more useful, is it
> >>> possible to make any of this the default? Can a system where it is
> >>> advantageous be autodetected?
> >>>
> >>> Andrew.
> >