lkml.org 
[lkml]   [2017]   [Dec]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [f2fs-dev] [PATCH 1/2] f2fs: pass down write hints to block layer for bufferd write
Hi Jaegeuk,

On 12/28/2017 12:26 PM, Jaegeuk Kim wrote:
> On 12/23, Chao Yu wrote:
>> On 2017/12/15 10:06, Jaegeuk Kim wrote:
>>> On 12/14, Hyunchul Lee wrote:
>>>> Hi Jaegeuk,
>>>>
>>>> I need your comment about the fs_iohint mount option.
>>>>
>>>> a) w/o fs_iohint, propagate user hints to low layer.
>>>> b) w/ fs_iohint, ignore user hints, and use hints which is generated
>>>> with F2FS.
>>>>
>>>> Chao suggests this option. because user hints are more accurate than
>>>> file system.
>>>>
>>>> This is resonable, But I have some concerns about this option.
>>>> The first thing is that blocks of a segments have different hints. This
>>>> could make GC less effective.
>>>> The second is that the separation between LIFE_MEDIUM and LIFE_LONG is
>>>> really needed. I think that difference between them is a little ambigous
>>>> for users, and LIFE_SHORT and LIFE_EXTREME is converted to different
>>>> hints by F2FS.
>>>
>>> I think what we really can do would assign many user hints to our 3 DATA
>>> logs likewise rw_hint_to_seg_type(), since it's just hints for user data.
>>> Then, we can decide how to keep that as much as possible, since we have
>>> another filesystem metadata such as meta and nodes. In addition, I don't
>>> think we have to keep the original user-hints which makes F2FS logs be
>>> messed up.
>>>
>>> With that mind, I can think of the below cases. Especially, if user wants
>>> to keep their io_hints, we'd better recommend to use direct_io w/o fs_iohints.
>>
>>
>>
>>> In order to keep this policy, I think fs_iohints would be better to be a
>>> feature set by mkfs.f2fs and detected by sysfs entries for users.
>>>
>>> 1) w/ fs_iohints
>>>
>>> User F2FS Block
>>> -------------------------------------------------------------------
>>> Meta WRITE_LIFE_MEDIUM
>>> HOT_NODE WRITE_LIFE_NOTSET
>>> WARM_NODE -'
>>> COLD_NODE WRITE_LIFE_NONE
>>> ioctl(cold) COLD_DATA WRITE_LIFE_EXTREME
>>> extention list -' -'
>>> WRITE_LIFE_EXTREME -' -'
>>> WRITE_LIFE_SHORT HOT_DATA WRITE_LIFE_SHORT
>>>
>>> -- buffered_io
>>> WRITE_LIFE_NOT_SET WARM_DATA WRITE_LIFE_LONG
>>> WRITE_LIFE_NONE -' -'
>>> WRITE_LIFE_MEDIUM -' -'
>>> WRITE_LIFE_LONG -' -'
>>>
>>> -- direct_io (Not recommendable)
>>> WRITE_LIFE_NOT_SET WARM_DATA WRITE_LIFE_NOT_SET
>>> WRITE_LIFE_NONE -' WRITE_LIFE_NONE
>>> WRITE_LIFE_MEDIUM -' WRITE_LIFE_MEDIUM
>>> WRITE_LIFE_LONG -' WRITE_LIFE_LONG
>>
>> Agreed with above IO hint mapping rule.
>>
>>>
>>> 2) w/o fs_iohints
>>>
>>> User F2FS Block
>>> -------------------------------------------------------------------
>>> Meta -
>>> HOT_NODE -
>>> WARM_NODE -
>>> COLD_NODE -
>>> ioctl(cold) COLD_DATA -
>>> extention list -' -
>>>
>>> -- buffered_io
>>> WRITE_LIFE_EXTREME COLD_DATA -
>>> WRITE_LIFE_SHORT HOT_DATA -
>>> WRITE_LIFE_NOT_SET WARM_DATA -
>>> WRITE_LIFE_NONE -' -
>>> WRITE_LIFE_MEDIUM -' -
>>> WRITE_LIFE_LONG -' -
>>
>> Now we recommend direct_io if user wants to give IO hint for storage, I suspect
>> that user would suffer performance regression issue w/o buffered IO.
>>
>> Another problem is that, now, in Android, it will be very hard to prompt
>> application to migrate their IO pattern from buffered IO to direct IO, one
>> possible way is distinguishing user data lifetime from FWK, e.g. set
>> WRITE_LIFE_SHORT for cache file or tmp file, set WRITE_LIFE_EXTREME for media file.
>>
>> In order to support buffered_io, would it be better to change mapping as below?
>>
>> -- buffered_io
>> WRITE_LIFE_EXTREME COLD_DATA WRITE_LIFE_EXTREME
>> WRITE_LIFE_SHORT HOT_DATA WRITE_LIFE_SHORT
>> WRITE_LIFE_NOT_SET WARM_DATA WRITE_LIFE_NOT_SET
>> WRITE_LIFE_NONE -' -'
>> WRITE_LIFE_MEDIUM -' -'
>> WRITE_LIFE_LONG -' -'
>
> Agreed, and it makes more sense that we'd better keep the write hints on
> userdata given by applications.
>
> BTW, since we couldn't get any performance numbers with these, how about
> adding a mount option like "-o iohints=MODE" where MODE may be one of
> "fs-based", "user-based", and "off"?
>

"fs-based" equals "with fs_iohints", "user-based" equals "without fs_iohints"
+ Chao's suggest, and "off" means not passing down hints to block layer. right?

Thanks.

> Thanks,
>
>>
>> Thanks,
>>
>>>
>>> -- direct_io
>>> WRITE_LIFE_EXTREME COLD_DATA WRITE_LIFE_EXTREME
>>> WRITE_LIFE_SHORT HOT_DATA WRITE_LIFE_SHORT
>>> WRITE_LIFE_NOT_SET WARM_DATA WRITE_LIFE_NOT_SET
>>> WRITE_LIFE_NONE -' WRITE_LIFE_NONE
>>> WRITE_LIFE_MEDIUM -' WRITE_LIFE_MEDIUM
>>> WRITE_LIFE_LONG -' WRITE_LIFE_LONG
>>>
>>>
>>> Note that, I don't much care about how to manipulate streamid in nvme driver
>>> in terms of LIFE_NONE or LIFE_NOTSET, since other drivers can handle them
>>> in different ways. Taking a look at the definition, at least, we don't need
>>> to assume that those are same at all. For example, if we can expolit this in
>>> UFS driver, we can pass all the stream ids to the device as context ids.
>>>
>>> Thanks,
>>>
>>>>
>>>> Thanks.
>>>>
>>>> On 12/12/2017 11:45 AM, Chao Yu wrote:
>>>>> Hi Hyunchul,
>>>>>
>>>>> On 2017/12/12 10:15, Hyunchul Lee wrote:
>>>>>> Hi Chao,
>>>>>>
>>>>>> On 12/11/2017 10:15 PM, Chao Yu wrote:
>>>>>>> Hi Hyunchul,
>>>>>>>
>>>>>>> On 2017/12/1 16:28, Hyunchul Lee wrote:
>>>>>>>> Hi Chao,
>>>>>>>>
>>>>>>>> On 11/30/2017 04:06 PM, Chao Yu wrote:
>>>>>>>>> Hi Hyunchul,
>>>>>>>>>
>>>>>>>>> On 2017/11/28 8:23, Hyunchul Lee wrote:
>>>>>>>>>> From: Hyunchul Lee <cheol.lee@lge.com>
>>>>>>>>>>
>>>>>>>>>> This implements which hint is passed down to block layer
>>>>>>>>>> for datas from the specific segment type.
>>>>>>>>>>
>>>>>>>>>> segment type hints
>>>>>>>>>> ------------ -----
>>>>>>>>>> COLD_NODE & COLD_DATA WRITE_LIFE_EXTREME
>>>>>>>>>> WARM_DATA WRITE_LIFE_NONE
>>>>>>>>>> HOT_NODE & WARM_NODE WRITE_LIFE_LONG
>>>>>>>>>> HOT_DATA WRITE_LIFE_MEDIUM
>>>>>>>>>> META_DATA WRITE_LIFE_SHORT
>>>>>>>>>
>>>>>>>>> Just noticed, if our user do not give the hint via ioctl, f2fs can
>>>>>>>>> provider hint to lower layer according to hot/cold separation ability,
>>>>>>>>> it will be okay. But once user give his hint which may be more accurate
>>>>>>>>> than filesystem, hint converted by f2fs may be wrong.
>>>>>>>>>
>>>>>>>>> So what do you think of adding an option to control whether filesystem
>>>>>>>>> can convert hint user given?
>>>>>>>>>
>>>>>>>>
>>>>>>>> I think it is okay for LIFE_SHORT and LIFE_EXTREME. because they are
>>>>>>>> converted to different hints.
>>>>>>>
>>>>>>> What I mean is introducing a mount option, e.g. fs_iohint,
>>>>>>> a) w/o fs_iohint, propagate file/inode io_hint to low layer.
>>>>>>> b) w/ fs_iohint, ignore file/inode io_hint, use io_hint which is generated
>>>>>>> with filesystem's private rule.
>>>>>>>
>>>>>>
>>>>>> Okay, I will implement this option and send this patch again.
>>>>>
>>>>> Let's wait for Jaegeuk's comments first?
>>>>>
>>>>>>
>>>>>> Without fs_iohint, Even if data blocks are moved due to GC,
>>>>>> we should keep user hints. And if user hints are not given,
>>>>>> any hints are not passed down to block layer, right?
>>>>>
>>>>> Hmm.. that will be a problem, IMO, we can store last user's io_hint into inode
>>>>> layout, so later when we trigger GC, we can use the last io_hint in inode rather
>>>>> than giving no hint or fs' hint.
>>>>>
>>>>> I think it needs to discuss with original author of IO hint, what is the IO hint
>>>>> policy when filesystem move block by itself after inode has been released in system.
>>>>>
>>>>> Thanks,
>>>>>
>>>>>>
>>>>>> Thank you for comments.
>>>>>>
>>>>>>> Thanks,
>>>>>>>
>>>>>>>>
>>>>>>>> file hint segment type io hint
>>>>>>>> --------- ------------ -------
>>>>>>>> LIFE_SHORT HOT_DATA LIFE_MEDIUM
>>>>>>>> LIFE_MEDIUM WARM_DATA LIFE_NONE
>>>>>>>> LIFE_LONG WARM_DATA LIFE_NONE
>>>>>>>> LIFE_EXTREME COLD_DATA LIFE_EXTREME
>>>>>>>>
>>>>>>>> the problem is that LIFE_MEDIUM and LIFE_LONG are converted to
>>>>>>>> the same hint, LIFE_NONE. I am not sure that the seperation between
>>>>>>>> LIFE_MEDIUM and LIFE_LONG is really needed. Because I guess that the
>>>>>>>> difference between them is a little ambigous for users, and if WARM_DATA
>>>>>>>> segment has two different hints, it can makes GC non-efficient.
>>>>>>>>
>>>>>>>> I wonder your thought about this.
>>>>>>>>
>>>>>>>> Thanks.
>>>>>>>>
>>>>>>>>> Thanks,
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>> ------------------------------------------------------------------------------
>>>>>>>> Check out the vibrant tech community on one of the world's most
>>>>>>>> engaging tech sites, Slashdot.org! http://sdm.link/slashdot
>>>>>>>> _______________________________________________
>>>>>>>> Linux-f2fs-devel mailing list
>>>>>>>> Linux-f2fs-devel@lists.sourceforge.net
>>>>>>>> https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>> .
>>>>>>
>>>>>
>>>>>
>>>>> ------------------------------------------------------------------------------
>>>>> Check out the vibrant tech community on one of the world's most
>>>>> engaging tech sites, Slashdot.org! http://sdm.link/slashdot
>>>>> _______________________________________________
>>>>> Linux-f2fs-devel mailing list
>>>>> Linux-f2fs-devel@lists.sourceforge.net
>>>>> https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel
>>>>>
>>>
>>> .
>>>
>

\
 
 \ /
  Last update: 2017-12-28 06:06    [W:0.144 / U:0.064 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site