On Jan 6, 2011, at 4:57 PM, Brice Goglin wrote:
> Is HWLOC_MEMBIND_MIXED defined somewhere?
(sorry for the latency; I got distracted by other things)
Sorry -- Samuel and I talked about this one in IM. I assumed that Samuel would be adding it. :-)
Even though we don't currently need it anywhere, we need to provide the *possibility* for this corner case.
> I am not sure about "on next touch ... move and rebind" in the NEXTTOUCH
>> From what I see, on solaris, it's only supported when we bind to the
> entire machine. So there's no binding, only moving there.
> Linux doesn't support NEXTTOUCH yet. I think most implementations I have
> seen would change the binding immediately and only move the page
> (according to the binding) on next-touch.
> In short, my feeling is that it's more "bind but only lazily move pages
> on next touch" (compared to MIGRATE = "bind and move everything now",
> and BIND = "bind but don't move already faulted pages").
Hmm; I thought I said that, given the "On next touch of each page..." phrase:
HWLOC_MEMBIND_NEXTTOUCH = 5 /**< \brief On next touch of
* each page in the existing
* allocated memory, migrate
* (i.e., move and re-bind) it
* to the local NUMA node of
* the thread where the memory
* reference occurred.
* \hideinitializer */
So how's this:
As each page bound with this policy is touched, it is moved from its current location to the local NUMA node of the thread where the memory reference occurred (if it needs to be moved at all).
For corporate legal information go to: