Re: Opinions AMD64 vs EM64T vs Itanium II

On Sat, 2006-08-26 at 12:20 -0700, Jeffrey Siegal wrote:
> On Aug 26, 2006, at 00:12 , Arjan van de Ven wrote:
> 
> > Cons are
> > that your system de facto is NUMA and that you get all the problems
> > associated with that: you only get this gain if both the kernel and  
> > the
> > applications are NUMA aware. The kernel mostly is, applications...
> > mostly not.
> 
> What could or should the applications be doing?  I would think that  
> allocating pages being accessed by a process on CPU N to memory  
> attached to CPU N would be the job of the kernel, right?

Hi,

yes that bit is a kernel job for the simple case; it's not the case I
meant though: 
The big difference is this, and this only matters for multithreaded apps
or apps that share memory between processes like oracle: If you have a
cpu intensive application that shares memory and uses it across all
threads all the time, then your app is going to show "really bad" numa
behavior (what the effect of "bad numa behavior" is depends on how
costly that is on the hardware, different topic ;). It effectively is a
case that the memory doesn't have any logical local place to live, so
there is no "home node" for it.

Numa aware applications avoid this behavior. It's not always trivial to
make such applications numa aware, and may require massive
rearchitecting...

Greetings,
   Arjan van de Ven
-- 
if you want to mail me at work (you don't), use arjan (at) linux.intel.com