2020-11-03 08:40:51

by Daejun Park

[permalink] [raw]
Subject: [PATCH v2] f2fs: change write_hint for hot/warm nodes

In the fs-based mode of F2FS, the mapping of hot/warm node to
WRITE_LIFE_NOT_SET should be changed to WRITE_LIFE_SHORT.

As a result of analyzing the write pattern of f2fs using real workload,
hot/warm nodes have high update ratio close to hot data.[*]
However, F2FS passes write hints for hot/warm nodes as WRITE_LIFE_NOT_SET.

Furthermore, WRITE_LIFE_NOT_SET is a default value of write hint when it is
not provided from the file system.
In storage, write_hint is used to distinguish streams (e.g. NVMe).
So, the hot/warm node of F2FS is not distinguished from other write_hints,
which can make the wrong stream seperation.

* Liang, Yu, et al. "An empirical study of F2FS on mobile devices." 2017
IEEE 23rd International Conference on Embedded and Real-Time Computing
Systems and Applications (RTCSA).

Signed-off-by: Daejun Park <[email protected]>
---
v2: update documentation and comments
---
Documentation/filesystems/f2fs.rst | 4 ++--
fs/f2fs/segment.c | 6 +++---
2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/Documentation/filesystems/f2fs.rst b/Documentation/filesystems/f2fs.rst
index b8ee761c9922..afa3da7cfade 100644
--- a/Documentation/filesystems/f2fs.rst
+++ b/Documentation/filesystems/f2fs.rst
@@ -717,8 +717,8 @@ WRITE_LIFE_LONG " WRITE_LIFE_LONG
===================== ======================== ===================
User F2FS Block
===================== ======================== ===================
- META WRITE_LIFE_MEDIUM;
- HOT_NODE WRITE_LIFE_NOT_SET
+ META WRITE_LIFE_MEDIUM
+ HOT_NODE WRITE_LIFE_SHORT
WARM_NODE "
COLD_NODE WRITE_LIFE_NONE
ioctl(COLD) COLD_DATA WRITE_LIFE_EXTREME
diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
index 1596502f7375..577ab7516c6b 100644
--- a/fs/f2fs/segment.c
+++ b/fs/f2fs/segment.c
@@ -3160,8 +3160,8 @@ int f2fs_rw_hint_to_seg_type(enum rw_hint hint)
*
* User F2FS Block
* ---- ---- -----
- * META WRITE_LIFE_MEDIUM;
- * HOT_NODE WRITE_LIFE_NOT_SET
+ * META WRITE_LIFE_MEDIUM
+ * HOT_NODE WRITE_LIFE_SHORT
* WARM_NODE "
* COLD_NODE WRITE_LIFE_NONE
* ioctl(COLD) COLD_DATA WRITE_LIFE_EXTREME
@@ -3208,7 +3208,7 @@ enum rw_hint f2fs_io_type_to_rw_hint(struct f2fs_sb_info *sbi,
return WRITE_LIFE_EXTREME;
} else if (type == NODE) {
if (temp == WARM || temp == HOT)
- return WRITE_LIFE_NOT_SET;
+ return WRITE_LIFE_SHORT;
else if (temp == COLD)
return WRITE_LIFE_NONE;
} else if (type == META) {
--
2.25.1


2020-11-03 09:05:33

by Chao Yu

[permalink] [raw]
Subject: Re: [PATCH v2] f2fs: change write_hint for hot/warm nodes

On 2020/11/3 16:38, Daejun Park wrote:
> In the fs-based mode of F2FS, the mapping of hot/warm node to
> WRITE_LIFE_NOT_SET should be changed to WRITE_LIFE_SHORT.
>
> As a result of analyzing the write pattern of f2fs using real workload,
> hot/warm nodes have high update ratio close to hot data.[*]
> However, F2FS passes write hints for hot/warm nodes as WRITE_LIFE_NOT_SET.
>
> Furthermore, WRITE_LIFE_NOT_SET is a default value of write hint when it is
> not provided from the file system.
> In storage, write_hint is used to distinguish streams (e.g. NVMe).
> So, the hot/warm node of F2FS is not distinguished from other write_hints,
> which can make the wrong stream seperation.
>
> * Liang, Yu, et al. "An empirical study of F2FS on mobile devices." 2017
> IEEE 23rd International Conference on Embedded and Real-Time Computing
> Systems and Applications (RTCSA).
>
> Signed-off-by: Daejun Park <[email protected]>

Reviewed-by: Chao Yu <[email protected]>

Thanks,

2020-11-03 19:02:12

by Jaegeuk Kim

[permalink] [raw]
Subject: Re: [PATCH v2] f2fs: change write_hint for hot/warm nodes

On 11/03, Daejun Park wrote:
> In the fs-based mode of F2FS, the mapping of hot/warm node to
> WRITE_LIFE_NOT_SET should be changed to WRITE_LIFE_SHORT.
>
> As a result of analyzing the write pattern of f2fs using real workload,
> hot/warm nodes have high update ratio close to hot data.[*]
> However, F2FS passes write hints for hot/warm nodes as WRITE_LIFE_NOT_SET.

I prefer to keep it as is, since basically node blocks should be separatly
stored from data blocks in different erase blocks in order to match F2FS GC
and FTL GC units. And, we don't do IPU for node blocks which doesn't make sense
to say *update ratio*.

>
> Furthermore, WRITE_LIFE_NOT_SET is a default value of write hint when it is
> not provided from the file system.
> In storage, write_hint is used to distinguish streams (e.g. NVMe).
> So, the hot/warm node of F2FS is not distinguished from other write_hints,
> which can make the wrong stream seperation.
>
> * Liang, Yu, et al. "An empirical study of F2FS on mobile devices." 2017
> IEEE 23rd International Conference on Embedded and Real-Time Computing
> Systems and Applications (RTCSA).
>
> Signed-off-by: Daejun Park <[email protected]>
> ---
> v2: update documentation and comments
> ---
> Documentation/filesystems/f2fs.rst | 4 ++--
> fs/f2fs/segment.c | 6 +++---
> 2 files changed, 5 insertions(+), 5 deletions(-)
>
> diff --git a/Documentation/filesystems/f2fs.rst b/Documentation/filesystems/f2fs.rst
> index b8ee761c9922..afa3da7cfade 100644
> --- a/Documentation/filesystems/f2fs.rst
> +++ b/Documentation/filesystems/f2fs.rst
> @@ -717,8 +717,8 @@ WRITE_LIFE_LONG " WRITE_LIFE_LONG
> ===================== ======================== ===================
> User F2FS Block
> ===================== ======================== ===================
> - META WRITE_LIFE_MEDIUM;
> - HOT_NODE WRITE_LIFE_NOT_SET
> + META WRITE_LIFE_MEDIUM
> + HOT_NODE WRITE_LIFE_SHORT
> WARM_NODE "
> COLD_NODE WRITE_LIFE_NONE
> ioctl(COLD) COLD_DATA WRITE_LIFE_EXTREME
> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
> index 1596502f7375..577ab7516c6b 100644
> --- a/fs/f2fs/segment.c
> +++ b/fs/f2fs/segment.c
> @@ -3160,8 +3160,8 @@ int f2fs_rw_hint_to_seg_type(enum rw_hint hint)
> *
> * User F2FS Block
> * ---- ---- -----
> - * META WRITE_LIFE_MEDIUM;
> - * HOT_NODE WRITE_LIFE_NOT_SET
> + * META WRITE_LIFE_MEDIUM
> + * HOT_NODE WRITE_LIFE_SHORT
> * WARM_NODE "
> * COLD_NODE WRITE_LIFE_NONE
> * ioctl(COLD) COLD_DATA WRITE_LIFE_EXTREME
> @@ -3208,7 +3208,7 @@ enum rw_hint f2fs_io_type_to_rw_hint(struct f2fs_sb_info *sbi,
> return WRITE_LIFE_EXTREME;
> } else if (type == NODE) {
> if (temp == WARM || temp == HOT)
> - return WRITE_LIFE_NOT_SET;
> + return WRITE_LIFE_SHORT;
> else if (temp == COLD)
> return WRITE_LIFE_NONE;
> } else if (type == META) {
> --
> 2.25.1

2020-11-05 04:59:47

by Chao Yu

[permalink] [raw]
Subject: Re: [PATCH v2] f2fs: change write_hint for hot/warm nodes

On 2020/11/4 2:58, Jaegeuk Kim wrote:
> On 11/03, Daejun Park wrote:
>> In the fs-based mode of F2FS, the mapping of hot/warm node to
>> WRITE_LIFE_NOT_SET should be changed to WRITE_LIFE_SHORT.
>>
>> As a result of analyzing the write pattern of f2fs using real workload,
>> hot/warm nodes have high update ratio close to hot data.[*]
>> However, F2FS passes write hints for hot/warm nodes as WRITE_LIFE_NOT_SET.
>
> I prefer to keep it as is, since basically node blocks should be separatly
> stored from data blocks in different erase blocks in order to match F2FS GC

Correct, I missed that the change will break HOT_DATA type and {HOT,WARM}_NODE
type data separation in FTL.

> and FTL GC units. And, we don't do IPU for node blocks which doesn't make sense
> to say *update ratio*.

IMO, if fs discard is fast enough, both IPU and OPU will cause original blkaddr in
FTL being invalid, we can compare update ratio on {I,O}PU IOs inside FTL; otherwise,
we can't.

Thanks,

>
>>
>> Furthermore, WRITE_LIFE_NOT_SET is a default value of write hint when it is
>> not provided from the file system.
>> In storage, write_hint is used to distinguish streams (e.g. NVMe).
>> So, the hot/warm node of F2FS is not distinguished from other write_hints,
>> which can make the wrong stream seperation.
>>
>> * Liang, Yu, et al. "An empirical study of F2FS on mobile devices." 2017
>> IEEE 23rd International Conference on Embedded and Real-Time Computing
>> Systems and Applications (RTCSA).
>>
>> Signed-off-by: Daejun Park <[email protected]>
>> ---
>> v2: update documentation and comments
>> ---
>> Documentation/filesystems/f2fs.rst | 4 ++--
>> fs/f2fs/segment.c | 6 +++---
>> 2 files changed, 5 insertions(+), 5 deletions(-)
>>
>> diff --git a/Documentation/filesystems/f2fs.rst b/Documentation/filesystems/f2fs.rst
>> index b8ee761c9922..afa3da7cfade 100644
>> --- a/Documentation/filesystems/f2fs.rst
>> +++ b/Documentation/filesystems/f2fs.rst
>> @@ -717,8 +717,8 @@ WRITE_LIFE_LONG " WRITE_LIFE_LONG
>> ===================== ======================== ===================
>> User F2FS Block
>> ===================== ======================== ===================
>> - META WRITE_LIFE_MEDIUM;
>> - HOT_NODE WRITE_LIFE_NOT_SET
>> + META WRITE_LIFE_MEDIUM
>> + HOT_NODE WRITE_LIFE_SHORT
>> WARM_NODE "
>> COLD_NODE WRITE_LIFE_NONE
>> ioctl(COLD) COLD_DATA WRITE_LIFE_EXTREME
>> diff --git a/fs/f2fs/segment.c b/fs/f2fs/segment.c
>> index 1596502f7375..577ab7516c6b 100644
>> --- a/fs/f2fs/segment.c
>> +++ b/fs/f2fs/segment.c
>> @@ -3160,8 +3160,8 @@ int f2fs_rw_hint_to_seg_type(enum rw_hint hint)
>> *
>> * User F2FS Block
>> * ---- ---- -----
>> - * META WRITE_LIFE_MEDIUM;
>> - * HOT_NODE WRITE_LIFE_NOT_SET
>> + * META WRITE_LIFE_MEDIUM
>> + * HOT_NODE WRITE_LIFE_SHORT
>> * WARM_NODE "
>> * COLD_NODE WRITE_LIFE_NONE
>> * ioctl(COLD) COLD_DATA WRITE_LIFE_EXTREME
>> @@ -3208,7 +3208,7 @@ enum rw_hint f2fs_io_type_to_rw_hint(struct f2fs_sb_info *sbi,
>> return WRITE_LIFE_EXTREME;
>> } else if (type == NODE) {
>> if (temp == WARM || temp == HOT)
>> - return WRITE_LIFE_NOT_SET;
>> + return WRITE_LIFE_SHORT;
>> else if (temp == COLD)
>> return WRITE_LIFE_NONE;
>> } else if (type == META) {
>> --
>> 2.25.1
> .
>