Re: [RFC PATCH 0/2] apply write hints to select the type of segments

[Hyunchul Lee]

On 11/18/2017 03:53 AM, Jaegeuk Kim wrote:
> ...
>>>>>>>>>>>>>>>>> From: Hyunchul Lee <cheol.lee@xxxxxxx>
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> Using write hints[1], applications can inform the life time of the data
>>>>>>>>>>>>>>>>> written to devices. and this[2] reported that the write hints patch
>>>>>>>>>>>>>>>>> decreased writes in NAND by 25%.
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> This hints help F2FS to determine the followings.
>>>>>>>>>>>>>>>>>   1) the segment types where the data will be written.
>>>>>>>>>>>>>>>>>   2) the hints that will be passed down to devices with the data of segments.
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> This patch set implements the first mapping from write hints to segment types
>>>>>>>>>>>>>>>>> as shown below.
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>   hints                     segment type
>>>>>>>>>>>>>>>>>   -----                     ------------
>>>>>>>>>>>>>>>>>   WRITE_LIFE_SHORT          CURSEG_COLD_DATA
>>>>>>>>>>>>>>>>>   WRITE_LIFE_EXTREME        CURSEG_HOT_DATA
>>>>>>>>>>>>>>>>>   others                    CURSEG_WARM_DATA
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> The F2FS poliy for hot/cold seperation has precedence over this hints, And
>>>>>>>>>>>>>>>>> hints are not applied in in-place update.
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Could we change to disable IPU if file/inode write hint is existing?
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> I am afraid that this makes side effects. for example, this could cause
>>>>>>>>>>>>>>> out-of-place updates even when there are not enough free segments. 
>>>>>>>>>>>>>>> I can write the patch that handles these situations. But I wonder 
>>>>>>>>>>>>>>> that this is required, and I am not sure which IPU polices can be disabled.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Oh, As I replied in another thread, I think IPU just affects filesystem
>>>>>>>>>>>>>> hot/cold separating, rather than this feature. So I think it will be okay
>>>>>>>>>>>>>> to not consider it.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> Before the second mapping is implemented, write hints are not passed down
>>>>>>>>>>>>>>>>> to devices. Because it is better that the data of a segment have the same 
>>>>>>>>>>>>>>>>> hint.
>>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>>> [1]: c75b1d9421f80f4143e389d2d50ddfc8a28c8c35
>>>>>>>>>>>>>>>>> [2]: https://lwn.net/Articles/726477/
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>> Could you write a patch to support passing write hint to block layer for
>>>>>>>>>>>>>>>> buffered writes as below commit:
>>>>>>>>>>>>>>>> 0127251c45ae ("ext4: add support for passing in write hints for buffered writes")
>>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Sure I will. I wrote it already ;)
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Cool, ;)
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> I think that datas from the same segment should be passed down with the same
>>>>>>>>>>>>>>> hint, and the following mapping is reasonable. I wonder what is your opinion
>>>>>>>>>>>>>>> about it.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>   segment type               hints
>>>>>>>>>>>>>>>   ------------               -----
>>>>>>>>>>>>>>>   CURSEG_COLD_DATA           WRITE_LIFE_EXTREME
>>>>>>>>>>>>>>>   CURSEG_HOT_DATA            WRITE_LIFE_SHORT
>>>>>>>>>>>>>>>   CURSEG_COLD_NODE           WRITE_LIFE_NORMAL
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> We have WRITE_LIFE_LONG defined rather than WRITE_LIFE_NORMAL in fs.h?
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>>   CURSEG_HOT_NODE            WRITE_LIFE_MEDIUM
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> As I know, in scenario of cell phone, data of meta_inode is hottest, then hot
>>>>>>>>>>>>>> data, warm node, and cold node should be coldest. So I suggested we can define
>>>>>>>>>>>>>> as below:
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> META_DATA			WRITE_LIFE_SHORT
>>>>>>>>>>>>>> HOT_DATA & WARM_NODE		WRITE_LIFE_MEDIUM
>>>>>>>>>>>>>> HOT_NODE & WARM_DATA		WRITE_LIFE_LONG
>>>>>>>>>>>>>> COLD_NODE & COLD_DATA		WRITE_LIFE_EXTREME
>>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> I agree, But I am not sure that assigning the same hint to a node and data
>>>>>>>>>>>>> segment is good. Because NVMe is likely to write them in the same erase 
>>>>>>>>>>>>> block if they have the same hint.
>>>>>>>>>>>>
>>>>>>>>>>>> If we do not give the hint, they can still be written to the same erase block,
>>>>>>>>>>
>>>>>>>>>> I mean it's possible to write them to the same erase block. :)
>>>>>>>>>>
>>>>>>>>>>>> right? it will not be worse?
>>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> If the hint is not given, I think that they could be written to 
>>>>>>>>>>> the same erase block, or not. But if we give the same hint, they are written
>>>>>>>>>>> to the same block.
>>>>>>>>>>
>>>>>>>>>> IMO, Only if underlying device can support more hint type or opened channels,
>>>>>>>>>> and actual temperature of data segment and node segment is quite different, we
>>>>>>>>>> can separate them.
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Okay, If Jaegeuk Kim agrees with this, I will submit the patch that 
>>>>>>>>> implements your proposed mapping.
>>>>>>>>
>>>>>>>> How about this? We'd better to split data and node blocks as much as possible.
>>>>>>>>
>>>>>>>> segment type                    hints
>>>>>>>> ------------                    -----
>>>>>>>> COLD_NODE & COLD_DATA		WRITE_LIFE_NONE
>>>>>>>
>>>>>>> WRITE_LIFE_NONE means there is no hints about write life time.
>>>>>>>
>>>>>>> Shouldn't we define COLD_NODE & COLD_DATA as WRITE_LIFE_EXTERME?
>>>>>>
>>>>>> The assumption would be to split different types of blocks by flash firmware,
>>>>>> so I think we can use WRITE_LIFE_NONE as a type as well.
>>>>>>
>>>>>
>>>>> WRITE_LIFE_NONE means that no stream id is specified. It equals WRITE_LIFE_NOT_SET.
>>>>
>>>> Rgith, I just saw nvme implementation:
>>>>
>>>> nvme_assign_write_stream
>>>>
>>>> 	enum rw_hint streamid = req->write_hint;
>>>>
>>>> 	if (streamid == WRITE_LIFE_NOT_SET || streamid == WRITE_LIFE_NONE)
>>>> 		streamid = 0;
>>>> 	else {
>>>> 		streamid--;
>>>> ...
>>>>
>>>>> So I think that we can define WARM_DATA as WRITE_LIFE_NONE, and
>>>>> COLD_NODE & COLD_DATA as WRITE_LIFE_EXTREME.
>>>
>>> What's the point?
>>>
>>> segment type                 hints                streamid
>>> -------------                -----                -------
>>> COLD_NODE & COLD_DATA        WRITE_LIFE_NONE      0
>>> WARM_DATA                    WRITE_LIFE_EXTERME   4
>>> HOT_NODE & WARM_NODE         WRITE_LIFE_LONG      3
>>> HOT_DATA                     WRITE_LIFE_MEDIUM    2
>>> META_DATA                    WRITE_LIFE_SHORT     1
>>>
>>> So, I don't think something is wrong. Again, I don't care about its hotness
>>> given to the naming, but do care how to split different types of blocks with
>>> different stream ids. Exceptions would be giving _SHORT or _MEDIUM which are
>>> likely to be latency-critical, since I guess firmware may be able to store them
>>> into SLC buffer.
>>>
>>> Am I missing that _NONE has another meaning?
>>>
>>
>> What I am worried about is that datas with no hint have WRITE_LIFE_NOT_SET(id 0).
>> If block devices have swap partitions and anothor file systems, cold datas could
>> be mixed with datas from that. Does this seems way too much?
> 
> That seems like how to distinguish write_hints across multiple partitions?
> 

What I intend is that because there could be another partitions and 
the default stream ID is 0, WRITE_LIFE_EXTREAM could be better than 
WRITE_LIFE_NONE for cold datas.

Thanks.

>> And I think that stream id 0 means disabling stream directives. 
>> Becasue NVME_RW_DTYPE_STREAMS is clear.
> 
> Then, I guess SSD FW will just handle 5 stream IDs including disabled 0.
> 
> Thanks,
>