Yes, I will submit the patch later.
On 05/09/2013 07:50, Jaegeuk Kim wrote:
Hi Jin,
2013-09-04 (수), 21:17 +0800, Jin Xu:
Hi Jaegeuk,
On 04/09/2013 17:40, Jaegeuk Kim wrote:
Hi Jin,
2013-09-04 (수), 07:59 +0800, Jin Xu:
Hi Jaegeuk,
On 03/09/2013 08:45, Jaegeuk Kim wrote:
Hi Jin,
[...]
It seems that we can obtain the performance gain just by setting the
MAX_VICTIM_SEARCH to 4096, for example.
So, how about just adding an ending criteria like below?
I agree that we could get the performance improvement by simply
enlarging the MAX_VICTIM_SEARCH to 4096, but I am concerning the
scalability a little bit. Because it might always searching the whole
bitmap in some cases, for example, when dirty segments is 4000 and
total segments is 409600.
[snip]
[...]
if (p->max_search > MAX_VICTIM_SEARCH)
p->max_search = MAX_VICTIM_SEARCH;
The optimization does not apply to SSR mode. There has a reason.
As noticed in the test, when SSR selected the segments that have most
garbage blocks, then when gc is needed, all the dirty segments might
have very less garbage blocks, thus the gc overhead is high. This might
lead to performance degradation. So the patch does not change the
victim selection policy for SSR.
I think it doesn't care.
GC is only triggered during the direct node block allocation.
What it means that we need to consider the number of GC triggers where
the GC triggers more frequently during the normal data allocation than
the node block allocation.
So, I think it would not degrade performance significatly.
BTW, could you show some numbers for this?
Or could you test what I suggested?
Thanks,
I re-ran the test and got the following result:
---------------------------------------
2GB SDHC
create 52023 files of size 32768 bytes
random re-write 100000 records of 4KB
---------------------------------------
| file creation (s) | rewrite time (s) | gc count | gc garbage blocks |
no patch 341 4227 1174 174840
patched 296 2995 634 109314
patched (KIM) 324 2958 645 106682
In this test, it does not show the minor performance degradation caused
by applying the patch to SSR mode. Instead, the performance is a little
better with what you suggested.
I agree that the performance degradation would not be significant even
it does degrade. I ever saw the minor degradation in some workloads, but
I didn't save the data.
So, I agree that we can apply the patch to SSR mode as well.
And do you still have concerns about the formula for calculating the #
of search?
Thank you for the test. :)
What I've concerned is that, if it is really important to get a victim
more accurately for the performance as you described, it doesn't need to
calculate the number of searches IMO. Just let's select nr_dirty. Why
not?
Only the thing that we should consider is to handle the case where the
nr_dirty is too large.
For this, we can just limit the # of searches to avoid performance
degradation.
Still actually, I'm not convincing the effectiveness of your formula.
If possible, could you show it with numbers?
It's not easy to prove the effectiveness of the formula. It's just for
eliminating my concern on the scalability of searching. Since it does
not matter much for the performance improvement, we can put it aside
and choose the simpler method as you suggested.
So, should I revise the patch based on what you suggested or will
you take care of it?
Could you make a patch with your performance description and sumbit it
again?
Thanks a lot,
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
