svn commit: r250543 - user/attilio/jeff-numa/sys/vm

[Attilio Rao]

On Mon, May 13, 2013 at 8:45 PM, John Baldwin <jhb at freebsd.org> wrote:
> On 5/13/13 2:23 PM, Attilio Rao wrote:
>> On Mon, May 13, 2013 at 7:51 PM, John Baldwin <jhb at freebsd.org> wrote:
>>> On 5/11/13 9:53 PM, Attilio Rao wrote:
>>>> Author: attilio
>>>> Date: Sun May 12 01:53:03 2013
>>>> New Revision: 250543
>>>> URL: http://svnweb.freebsd.org/changeset/base/250543
>>>>
>>>> Log:
>>>>   Scan all the domains when trying to allocate from freepages.
>>>>   Some domains may be empty but others may still have some available pages
>>>>   satisfying the allocations.
>>>
>>> It seems like you have nested loops here so you are doing N^2 scans
>>> rather than N?
>>
>> If you want to say so. In most cases N=2 and M=4. However also
>> consider the less overhead caming from all the adjustement of the
>> flind because of now split DOMAINS-NFREELISTS concepts.
>
> *sigh*
>
> No.  Read the code.

I wrote the code so I certainly read it.

> vm_phys_alloc_pages() does 'for i = 0; i <
> vm_ndomains; i++' around calls to vm_phys_alloc_pages_domain() which
> internally has its _own_ 'for i = 0; i < vm_ndomains; i++'.  Same with
> vm_phys_alloc_freelist_pages().  You don't need it in both places.

It is not true.
vm_phys_alloc_domain_pages() (if this is what you are speaking about)
doesn't have internally i < vm_ndomains.

> Also, vm_phys_alloc_pages() should now be able to just call
> vm_phys_alloc_freelist_pages() and just be a simple loop that only
> iterates on flind as it used to do before all the NUMA changes.

I don't mind either way. This means I should do a version that accepts
domain for vm_phys_alloc_freelist_pages(). I just thought this version
is smaller.

Attilio


--
Peace can only be achieved by understanding - A. Einstein