Re: Migrate pages from a ccNUMA node to another - patch

[Zoltan Menyhart]

Hirokazu Takahashi wrote:

> I guess aruguments src_node, mm and pte would be redundant since
> they can be looked up from old_p with the reverse mapping scheme.

In my version 0.2, I can do with only the following arguments:
 *		node:	Destination NUMA node
 *		mm:	-> victim "mm_struct"
 *		pte:	-> PTE of the page to be moved
(If I have "mm" at hand, why not to use it ? Why not to avoid fetching the r-map
page struct ?)

> >Notes: "pte" can be NULL if I do not know it apriori
> >       I cannot release "mm->page_table_lock" otherwise I have to re-scan the "mm->pgd".
> 
> Re-schan plicy would be much better since migrating pages is heavy work.
> I don't think that holding mm->page_table_lock for long time would be
> good idea.

Re-scanning is "cache killer", at least on IA64 with huge user memory size.
I have more than 512 Mbytes user memory and its PTEs do not fit into the L2 cache.

In my current design, I have the outer loops: PGD, PMD and PTE walking; and once
I find a valid PTE, I check it against the list of max. 2048 physical addresses as
the inner loop.
I reversed them: walking through the list of max. 2048 physical addresses as outer
loop and the PGD - PMD - PTE scans as inner loops resulted in 4 to 5 times slower
migration.

> How do you think about following algorism:
>   1. get mm->page_table_lock
>   2. chose some pages.
>   3. release mm->page_table_lock
>   4. call remap_onepage() against each page.
>   5. goto step1 if there remain pages to be migrated.

I want to move the most frequently used pages - at least with the HW assisted
hot page detection.
I take "mm->page_table_lock", I nuke the PTE. We've got a good chance that the CPU
using the page observes a page fault almost immediately. It enters the page fault
handler and gets blocked by "mm->page_table_lock". If I released the lock, the CPU
could continue and realize that there is nothing to do, the page fault has already
been repaired. In the mean time, it is me who wait for "mm->page_table_lock".
At worst this scenario happens 2048 times.
If I keep the lock, the victim CPU enters only once the page fault handler.

I think what we should do is to "pull in" pages in to a node rather than than
"pushing them out" for two reasons:
- the recipient CPU executes the migration instead of busy waiting for the lock
- there is chance that the recipient CPU will find the migrated data useful
  in its cache

Regards,

Zoltán Menyhárt
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/