From: Paul McKenney <Paul.McKenney@us...> - 2001-10-05 23:54:44
Here are the notes I took from the lse-tech presentation on simple-binding
NUMA APIs. If I missed anything, please let me know! I will be updating
the document over the next few days. I did not catch all the names, so
if you to be credited with your question
1. Make it clear that CPU/node bitmasks are always in "virtual" space.
2. Need way of binding already-running processes/tasks/threads.
[Gerrit: can we unify the self-bind and other-bind APIs?]
3. How to handle more than 64 CPUs? [Paul Jackson: incorporate
abstract-binding API, use simple binding for groups of up to
64 CPUs identified via the abstract-binding API.]
4. How to get capability for >64 CPUs on machines that need it without
impacting machines that don't need it?
5. Specify behavior of default case where there are no launch policies
6. ESRCH vs. ENOENT???
7. When doing a bindtonode(), what happens to memory already faulted in?
(Currently, just stays put, but should we move it? If we just leave
it, and the app wants it moved, how does it do this? madvise()?
Or do we just say that apps who care should plan ahead?)
8. Should the "restrict" arguments be explicit (as currently documented)
or implicit (e.g., a syscall to register a restriction for the
9. [Matt Dobson] Could we use a hierarchical notation to handle
patterned requests for larger groups of processors? For example,
one could imagine splitting the bitmap into two 32-bit sections
that covered two levels of a hierarchy, so that, for example,
0xffffffff0000000f would denote the first four CPUs out of each
group of 32. Thoughts?
10. [Kimio Suganuma] Do we really want to prohibit interleaving numbers
of CPUs from different nodes? This requires that the OS know the
maximum number of CPUs on a given node at boot time. This information
is not currently provided.
11. [Anton Blanchard] The hotplug patch removes the concept of
CPU numbers. How to reconcile NUMA APIs and the hotplug patch?
Get latest updates about Open Source Projects, Conferences and News.