Re: [PATCH 0/2] swap: improve swap I/O rate

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

[...]

>> [missing patch #3]
>> I tried to get a similar patch working for swap out in shrink_page_list. And
>> it worked in functional terms, but the additional mergin was negligible.
>
>
> I think we have already done it.
> Look at shrink_mem_cgroup_zone which ends up calling shrink_page_list so we already have applied
> I/O plugging.
>

I saw that code and it is part of the kernel I used to test my patches.
But despite that code and my additional experiments of plug/unplug in 
shrink_page_list the effective I/O size of swap write stays at almost 4k.

Thereby so far I can tell you that the plugs in shrink_page_list and 
shrink_mem_cgroup_zone aren't sufficient - at least for my case.
You saw the blocktrace summaries in my first mail, an excerpt of a write 
submission stream looks like that:

  94,4   10      465     0.023520923   116  A   W 28868648 + 8 <- (94,5) 
28868456
  94,5   10      466     0.023521173   116  Q   W 28868648 + 8 [kswapd0]
  94,5   10      467     0.023522048   116  G   W 28868648 + 8 [kswapd0]
  94,5   10      468     0.023522235   116  P   N [kswapd0]
  94,5   10      469     0.023759892   116  I   W 28868648 + 8 ( 237844) 
[kswapd0]
  94,5   10      470     0.023760079   116  U   N [kswapd0] 1
  94,5   10      471     0.023760360   116  D   W 28868648 + 8 ( 468) 
[kswapd0]
  94,4   10      472     0.023891235   116  A   W 28868656 + 8 <- (94,5) 
28868464
  94,5   10      473     0.023891454   116  Q   W 28868656 + 8 [kswapd0]
  94,5   10      474     0.023892110   116  G   W 28868656 + 8 [kswapd0]
  94,5   10      475     0.023944610   116  I   W 28868656 + 8 ( 52500) 
[kswapd0]
  94,5   10      476     0.023944735   116  U   N [kswapd0] 1
  94,5   10      477     0.023944892   116  D   W 28868656 + 8 ( 282) 
[kswapd0]
  94,5   16       19     0.024023192 16033  C   W 28868648 + 8 ( 262832) [0]
  94,5   24       37     0.024196752 14526  C   W 28868656 + 8 ( 251860) [0]
[...]

But we can split this discussion from my other two patches and I would 
be happy to provide my test environment for further tests if there are 
new suggestions/patches/...

>> Maybe the cond_resched triggers much mor often than I expected, I'm open for
>> suggestions regarding improving the pagout I/O sizes as well.
>
>
> We could enhance write out by batch like ext4_bio_write_page.
>

Do you mean the changes brought by "bd2d0210 ext4: use bio layer instead 
of buffer layer in mpage_da_submit_io" ?



-- 

GrA 1/4 sse / regards, Christian Ehrhardt
IBM Linux Technology Center, System z Linux Performance

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>