Re: Performance regression in scsi sequential throughput (iozone) due to "e084b - page-allocator: preserve PFN ordering when __GFP_COLD is set"

[Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>]

Christian Ehrhardt wrote:
> Mel Gorman wrote:
>> On Mon, Feb 15, 2010 at 04:46:53PM +0100, Christian Ehrhardt wrote:
> [...]
>>> The differences in asm are pretty much the same, as before rmqueue_bulk was already inlined the actually intended change to its parameters was negligible.
>>> I wondered if it would be important if that is a constant value (-1) or if the reason was caused by that shift. So I tried:
>>>
>>>  23 @@ -965,7 +965,7 @@
>>>  24                 set_page_private(page, migratetype);
>>>  25                 list = &page->lru;
>>>  26         }
>>>  27 -       __mod_zone_page_state(zone, NR_FREE_PAGES, -(i << order));
>>>  28 +       __mod_zone_page_state(zone, NR_FREE_PAGES, -i);
>>>  29         spin_unlock(&zone->lock);
>>>  30         return i;
>>>  31  }
>>>
> [...]
>> It "fixes" it only by not calling direct reclaim when it should :(
> 
> yeah as we both realized -1 was not right so it was more a crazy workaround :-)
> 
> Anyway after that being a dead end again I dug even deeper into the details of direct_reclaim - I think we can agree that out of the counters we already know the race between try_to_free making progress and get_page not getting a page causing the congestion_wait is source of the issue.
> 
> So what I tried to bring some more light into all that was extending my perf counters to track a few more details in direct_reclaim.
> Two patches are attached and apply after the other three already available in that thread.
> The intention is
> a) to track the time
>  a1) spent in try_to_free_pages
>  a2) consumed after try_to_free_pages until get_page_from_freelist
>  a3) spent in get_pages_from_freelist
> b1) after seeing that order!=0 -> drain_all_pages I wondered if that might differ even all calls look like they have zero
> b2) tracking the average amount of pages freed by try_to_free_pages for fast path and slow path (progres&!page)
> 
> Naming convention (otherwise it might get confusing here)
> Good case - the scenario e.g. with e084b and 5f8dcc21 reverted resulting in high throughput and a low ratio of direct_reclaim running into progress&!page
> Bad case - the scenario e.g. on a clean 2.6.32
> Fast path - direct reclaim calls that did not run into progress&!page
> Slow path - direct reclaim calls that ran into progress&!page ending up in a long congestion_wait and therefore called "slow" path
> 
> Mini summary of what we had before in huge tables:
>             fast path   slow path
> GOOD CASE      ~98%       ~1-3%
> BAD CASE       ~70%        ~30%
> -> leading to throughput impact of e.g. 600 mb/s with 16 iozone threads (worse with even more threads)
> 
> Out of the numbers I got the following things might help us to create a new approach to a solution.
> The timings showed that that the so called slow case is actually much faster passing though direct_reclaim in bad case.
> 
> GOOD CASE                                        duration
> a1 Fast-avg-duration_pre_ttf_2_post_ttf            164099
> a2 Fast-avg-duration_post_ttf_2_pre_get_page          459
> a3 Fast-avg-duration_pre_get_page_2_post_get_page     346
> a1 Slow-avg-duration_pre_ttf_2_post_ttf            127621
> a2 Slow-avg-duration_post_ttf_2_pre_get_page         1957
> a3 Slow-avg-duration_pre_get_page_2_post_get_page     256
> BAD CASE                                         duration   deviation to good case in %
> a1 Fast-avg-duration_pre_ttf_2_post_ttf            122921   -25.09%
> a2 Fast-avg-duration_post_ttf_2_pre_get_page          521   13.53%
> a3 Fast-avg-duration_pre_get_page_2_post_get_page     244   -29.55%
> a1 Slow-avg-duration_pre_ttf_2_post_ttf            109740   -14.01%
> a2 Slow-avg-duration_post_ttf_2_pre_get_page          250   -87.18%
> a3 Slow-avg-duration_pre_get_page_2_post_get_page     117   -54.16%
> 
> That means that in the bad case the execution is much faster. Especially in the case that eventually runs into the slow path try_to_free is 14% faster, more important the time between try_to_free and get_pages is 87%! faster => less than a fifth and finally get_page is 54% faster, but that is probably just failing in an almost empty list which is fast^^.
> 
> As I checked order which always was zero the time is not spent in drain_all_pages and the only other thing left might be cond_resched ?!
> Everything else are a few assignments so it can't be much else.
> But why would e.g. not running into schedule in cond_resched cause get_pages to not find anything - I don't know and I would expect it should be the other way around - the faster you get from free to get the more pages should be left.

THe reason here is probably the the fact that in the bad case a lot of 
processes are waiting on congestion_wait and are therefore not runnnable 
and that way not scheduled via cond_resched.

I'll test this theory today or tomorrow with cond_resched in 
direct_reclaim commented out and expect almost no difference.

> I thought the progress try_to_free_pages is doing might be important as well so I took numbers for that too.
> From those I see that the good case as well as the bad case has an average of 62 pages freed in fast path.
> But in slow path they start to differ - while the good case that is running only seldom in that path anyway frees an average of 1.46 pages (that might be the thing causing it not getting a page eventually) in the bad case it makes a progress of avg 37 pages even in slow path.
> 
> PAGES-FREED  fast path   slow path
> GOOD CASE      ~62       ~1.46
> BAD CASE       ~62       ~37

5f8dcc21 introduced per migrate type pcp lists, is it possible that we 
run in a scenario where try_to_free frees a lot of pages via, but of the 
wrong migrate type?
And afterwards get_page
At least try_to_free_page and it's called shrink_ functions is not 
migrate type aware while get_page and its subsequent buffered_rmqueue 
and rmqueue_bulk are - btw here comes patch e084b.

I only see buffered_rmqueue chose a specific pcp list based on migrate 
type, and a fallback to migrate_reserve - is that enough fallback, what 
if the reserve is empty too but a few other types would not and those 
other types are the ones filled by try_to_free?

I'll try to get a per migrate type #pages statistic after direct_reclaim 
reaches !page - maybe that can confirm some parts of my theory.

> Thinking of it as asking "how few pages do I have to free until I fall from fast to slow path" the kernels behave different it looks wrong but interesting.
> The good case only drops to slow path (!page) in case of ~1.46 pages freed while the bad case seems to enter that much earlier with even 37 pages freed.
> 
> As order is always 0 and get_page afaik about getting just "one" page I wonder where these 37 pages disappeared especially as in bad case it is much faster getting to get_pages after freeing those ~37 pages.
> 
> Comments and ideas welcome!
> 
> 

-- 

Grüsse / regards, Christian Ehrhardt
IBM Linux Technology Center, Open Virtualization