[f2fs-dev] f2fs compress performance problem

[f2fs-dev] f2fs compress performance problem

[changfengnan]

Hi:

I've been working on f2fs compression for a while, I'm confused on f2fs
compression performance, after a while reserch, 
I found some problem, maybe need some discuss.
I use AndroBench test performance on mobile, after enable compression, the
benchmark scores have dropped a lot.
Specifically:
1. 32M sequential read has dropped to 50% of original. Test case open file
with O_RDONLY|O_DIRECT, and set POSIX_FADV_RANDOM, the major resaon 
is disable readahed. For now,I didn't found any patch can improve this. 
2. 4K random read has dropped to 40% of original, after merge `f2fs:
compress: add compress_inode to cache compressed blocks`,
significant improvement in random read performance, up to 90% of original,
maybe more.
3. 32M sequential overwrite has dropped to 10% of original, after merge
`f2fs: compress: remove unneeded read when rewrite whole cluster`
up to 30% of original.
4. 4K random read has dropped to 1% of original, yes only 1% of original, I
found  open file with O_WRONLY|O_DSYNC|O_DIRECT is  an important reason, 
every time sync a compress inode need do checkpoint, after I remove
checkpoint on compress inode, up to 10% of original. And I think major
reason of this
is we need read whole cluster and rewrite it ,but I did't think of any
method to improve this.

I want to know is there any idea can help to improve this.
And I want to know do we have goal for the performance of compression, is it
possible to achieve the original performance?

Thanks.



_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

Re: [f2fs-dev] f2fs compress performance problem

[Jaegeuk Kim]

On 06/04, changfengnan@vivo.com wrote:
> Hi:
> 
> I've been working on f2fs compression for a while, I'm confused on f2fs
> compression performance, after a while reserch, 
> I found some problem, maybe need some discuss.
> I use AndroBench test performance on mobile, after enable compression, the
> benchmark scores have dropped a lot.
> Specifically:
> 1. 32M sequential read has dropped to 50% of original. Test case open file
> with O_RDONLY|O_DIRECT, and set POSIX_FADV_RANDOM, the major resaon 
> is disable readahed. For now,I didn't found any patch can improve this. 
> 2. 4K random read has dropped to 40% of original, after merge `f2fs:
> compress: add compress_inode to cache compressed blocks`,
> significant improvement in random read performance, up to 90% of original,
> maybe more.
> 3. 32M sequential overwrite has dropped to 10% of original, after merge
> `f2fs: compress: remove unneeded read when rewrite whole cluster`
> up to 30% of original.
> 4. 4K random read has dropped to 1% of original, yes only 1% of original, I
> found  open file with O_WRONLY|O_DSYNC|O_DIRECT is  an important reason, 
> every time sync a compress inode need do checkpoint, after I remove
> checkpoint on compress inode, up to 10% of original. And I think major
> reason of this
> is we need read whole cluster and rewrite it ,but I did't think of any
> method to improve this.
> 
> I want to know is there any idea can help to improve this.
> And I want to know do we have goal for the performance of compression, is it
> possible to achieve the original performance?

Could you please check compress_cache and extent_cache that can improve read
performance? Both were done quite recently.

> 
> Thanks.


_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

Re: [f2fs-dev] f2fs compress performance problem

[Fengnan Chang]

Yes, I had enable compress_cache and extent_cache, compress_cache indeed 
can improve random read performance, but couldn't improve other test 
case.And extent_cache was default enabled in my test.


Fix the description of the previous email:
  4. 4K random overwrite has dropped to 1% of original, yes only 1% of 
original, I found  open file with O_WRONLY|O_DSYNC|O_DIRECT is an 
important reason, every time sync a compress inode need do checkpoint, 
after I remove checkpoint on compress inode, up to 10% of original. If I 
set write size equal to cluster size,  it can up to 35% of original. 
And I think major reason of this is we need read whole cluster and 
rewrite it , I've been trying to get around this restriction recently, 
but haven't made any progress yet.

Thanks.

On 2021/7/2 1:06, Jaegeuk Kim wrote:
> On 06/04, changfengnan@vivo.com wrote:
>> Hi:
>>
>> I've been working on f2fs compression for a while, I'm confused on f2fs
>> compression performance, after a while reserch,
>> I found some problem, maybe need some discuss.
>> I use AndroBench test performance on mobile, after enable compression, the
>> benchmark scores have dropped a lot.
>> Specifically:
>> 1. 32M sequential read has dropped to 50% of original. Test case open file
>> with O_RDONLY|O_DIRECT, and set POSIX_FADV_RANDOM, the major resaon
>> is disable readahed. For now,I didn't found any patch can improve this.
>> 2. 4K random read has dropped to 40% of original, after merge `f2fs:
>> compress: add compress_inode to cache compressed blocks`,
>> significant improvement in random read performance, up to 90% of original,
>> maybe more.
>> 3. 32M sequential overwrite has dropped to 10% of original, after merge
>> `f2fs: compress: remove unneeded read when rewrite whole cluster`
>> up to 30% of original.
>> 4. 4K random read has dropped to 1% of original, yes only 1% of original, I
>> found  open file with O_WRONLY|O_DSYNC|O_DIRECT is  an important reason,
>> every time sync a compress inode need do checkpoint, after I remove
>> checkpoint on compress inode, up to 10% of original. And I think major
>> reason of this
>> is we need read whole cluster and rewrite it ,but I did't think of any
>> method to improve this.
>>
>> I want to know is there any idea can help to improve this.
>> And I want to know do we have goal for the performance of compression, is it
>> possible to achieve the original performance?
> 
> Could you please check compress_cache and extent_cache that can improve read
> performance? Both were done quite recently.
> 
>>
>> Thanks.
> 


_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

Re: [f2fs-dev] f2fs compress performance problem

[Jaegeuk Kim]

On 07/02, Fengnan Chang wrote:
> Yes, I had enable compress_cache and extent_cache, compress_cache indeed can
> improve random read performance, but couldn't improve other test case.And
> extent_cache was default enabled in my test.

Sorry, we enabled extent_cache for RO partition only.

static inline bool f2fs_may_extent_tree(struct inode *inode)
{
	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);

	if (!test_opt(sbi, EXTENT_CACHE) ||
			is_inode_flag_set(inode, FI_NO_EXTENT) ||
			(is_inode_flag_set(inode, FI_COMPRESSED_FILE) &&
			 !f2fs_sb_has_readonly(sbi)))
		return false;
...


> 
> 
> Fix the description of the previous email:
>  4. 4K random overwrite has dropped to 1% of original, yes only 1% of
> original, I found  open file with O_WRONLY|O_DSYNC|O_DIRECT is an important
> reason, every time sync a compress inode need do checkpoint, after I remove
> checkpoint on compress inode, up to 10% of original. If I set write size
> equal to cluster size,  it can up to 35% of original. And I think major
> reason of this is we need read whole cluster and rewrite it , I've been
> trying to get around this restriction recently, but haven't made any
> progress yet.
> 
> Thanks.
> 
> On 2021/7/2 1:06, Jaegeuk Kim wrote:
> > On 06/04, changfengnan@vivo.com wrote:
> > > Hi:
> > > 
> > > I've been working on f2fs compression for a while, I'm confused on f2fs
> > > compression performance, after a while reserch,
> > > I found some problem, maybe need some discuss.
> > > I use AndroBench test performance on mobile, after enable compression, the
> > > benchmark scores have dropped a lot.
> > > Specifically:
> > > 1. 32M sequential read has dropped to 50% of original. Test case open file
> > > with O_RDONLY|O_DIRECT, and set POSIX_FADV_RANDOM, the major resaon
> > > is disable readahed. For now,I didn't found any patch can improve this.
> > > 2. 4K random read has dropped to 40% of original, after merge `f2fs:
> > > compress: add compress_inode to cache compressed blocks`,
> > > significant improvement in random read performance, up to 90% of original,
> > > maybe more.
> > > 3. 32M sequential overwrite has dropped to 10% of original, after merge
> > > `f2fs: compress: remove unneeded read when rewrite whole cluster`
> > > up to 30% of original.
> > > 4. 4K random read has dropped to 1% of original, yes only 1% of original, I
> > > found  open file with O_WRONLY|O_DSYNC|O_DIRECT is  an important reason,
> > > every time sync a compress inode need do checkpoint, after I remove
> > > checkpoint on compress inode, up to 10% of original. And I think major
> > > reason of this
> > > is we need read whole cluster and rewrite it ,but I did't think of any
> > > method to improve this.
> > > 
> > > I want to know is there any idea can help to improve this.
> > > And I want to know do we have goal for the performance of compression, is it
> > > possible to achieve the original performance?
> > 
> > Could you please check compress_cache and extent_cache that can improve read
> > performance? Both were done quite recently.
> > 
> > > 
> > > Thanks.
> > 


_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

Re: [f2fs-dev] f2fs compress performance problem

[Fengnan Chang]

I did more test on random overwrite, I set write size equal to cluster 
size, make sure always write one cluster, no cross cluster write and 
merge this patch 
https://lore.kernel.org/linux-f2fs-devel/6bd41278-0de8-5119-6857-0478c641cf71@vivo.com/T/#t 
,
it can up to 45% of original. This would seem to indicate that there are 
reasons other than the need for read whole cluster and rewrite it.
	


On 2021/7/2 14:24, Jaegeuk Kim wrote:
> On 07/02, Fengnan Chang wrote:
>> Yes, I had enable compress_cache and extent_cache, compress_cache indeed can
>> improve random read performance, but couldn't improve other test case.And
>> extent_cache was default enabled in my test.
> 
> Sorry, we enabled extent_cache for RO partition only.
> 
> static inline bool f2fs_may_extent_tree(struct inode *inode)
> {
> 	struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> 
> 	if (!test_opt(sbi, EXTENT_CACHE) ||
> 			is_inode_flag_set(inode, FI_NO_EXTENT) ||
> 			(is_inode_flag_set(inode, FI_COMPRESSED_FILE) &&
> 			 !f2fs_sb_has_readonly(sbi)))
> 		return false;
> ...
> 
> 
>>
>>
>> Fix the description of the previous email:
>>   4. 4K random overwrite has dropped to 1% of original, yes only 1% of
>> original, I found  open file with O_WRONLY|O_DSYNC|O_DIRECT is an important
>> reason, every time sync a compress inode need do checkpoint, after I remove
>> checkpoint on compress inode, up to 10% of original. If I set write size
>> equal to cluster size,  it can up to 35% of original. And I think major
>> reason of this is we need read whole cluster and rewrite it , I've been
>> trying to get around this restriction recently, but haven't made any
>> progress yet.
>>
>> Thanks.
>>
>> On 2021/7/2 1:06, Jaegeuk Kim wrote:
>>> On 06/04, changfengnan@vivo.com wrote:
>>>> Hi:
>>>>
>>>> I've been working on f2fs compression for a while, I'm confused on f2fs
>>>> compression performance, after a while reserch,
>>>> I found some problem, maybe need some discuss.
>>>> I use AndroBench test performance on mobile, after enable compression, the
>>>> benchmark scores have dropped a lot.
>>>> Specifically:
>>>> 1. 32M sequential read has dropped to 50% of original. Test case open file
>>>> with O_RDONLY|O_DIRECT, and set POSIX_FADV_RANDOM, the major resaon
>>>> is disable readahed. For now,I didn't found any patch can improve this.
>>>> 2. 4K random read has dropped to 40% of original, after merge `f2fs:
>>>> compress: add compress_inode to cache compressed blocks`,
>>>> significant improvement in random read performance, up to 90% of original,
>>>> maybe more.
>>>> 3. 32M sequential overwrite has dropped to 10% of original, after merge
>>>> `f2fs: compress: remove unneeded read when rewrite whole cluster`
>>>> up to 30% of original.
>>>> 4. 4K random read has dropped to 1% of original, yes only 1% of original, I
>>>> found  open file with O_WRONLY|O_DSYNC|O_DIRECT is  an important reason,
>>>> every time sync a compress inode need do checkpoint, after I remove
>>>> checkpoint on compress inode, up to 10% of original. And I think major
>>>> reason of this
>>>> is we need read whole cluster and rewrite it ,but I did't think of any
>>>> method to improve this.
>>>>
>>>> I want to know is there any idea can help to improve this.
>>>> And I want to know do we have goal for the performance of compression, is it
>>>> possible to achieve the original performance?
>>>
>>> Could you please check compress_cache and extent_cache that can improve read
>>> performance? Both were done quite recently.
>>>
>>>>
>>>> Thanks.
>>>
> 


_______________________________________________
Linux-f2fs-devel mailing list
Linux-f2fs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel