From: Jin Xu <[email protected]>
This patch improves the foreground gc efficiency by optimizing the
victim selection policy. With this optimization, the random re-write
performance could increase up to 20%.
For f2fs, foreground gc happens when disk is lack of free spaces,
it selects dirty segments and moves valid blocks around for making
more space available. The gc cost of a segment is determined by the
valid blocks in the segment. The less the valid blocks, the higher
the efficiency. The ideal victim segment is the one that has the
most garbage blocks.
Currently, it searches up to 20 dirty segments for a victim segment.
The selected victim is not likely the best victim for gc when there
are much more dirty segments. Why not searching more dirty segments
for a better victim? The cost of searching dirty segments is
negligible in comparison to moving blocks.
In this patch, it does not search a constant number of dirty segments
anymore, instead it calculates the number based on the total segments,
dirty segments and a threshold. Following is the pseudo formula.
,-- nr_dirty_segments, if total_segments < threshold
(# of search) = |
`-- (nr_dirty_segments * threshold) / total_segments,
Otherwise
The test case is simple. It creates as many files until the disk full.
The size for each file is 32KB. Then it writes as many as 100000
records of 4KB size to random offsets of random files in sync mode.
The testing was done on a 2GB partition of a SDHC card. Let's see the
test result of f2fs without and with the patch.
----without the patch----
created 52023 files of size 32768 bytes in 341 seconds
finished 100000 loops in 4204 seconds
(internally, 722 gc were done, 115939 garbage blocks were reclaimed)
----with the patch----
created 52023 files of size 32768 bytes in 326 seconds
finished 100000 loops in 3067 seconds
(internally, 1255 gc were done, 181101 garbage blocks were reclaimed)
It's obvious that, with the patch, f2fs finishes the test in 20+% less
time than without the patch.
Since the performance improvement is related to gc, it might not be so
obvious for other tests that do not trigger gc as often as this one (
This is because f2fs selects dirty segments for SSR use most of the
time when free space is in shortage). The well-known iozone test tool
was not used for benchmarking the patch becuase it seems do not have
a test case that performs random re-write on a full disk.
Signed-off-by: Jin Xu <[email protected]>
---
fs/f2fs/gc.c | 27 ++++++++++++++++++++++++++-
fs/f2fs/gc.h | 4 +++-
fs/f2fs/segment.h | 1 +
3 files changed, 30 insertions(+), 2 deletions(-)
diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
index 35f9b1a..4e045e6 100644
--- a/fs/f2fs/gc.c
+++ b/fs/f2fs/gc.c
@@ -138,10 +138,12 @@ static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
if (p->alloc_mode == SSR) {
p->gc_mode = GC_GREEDY;
p->dirty_segmap = dirty_i->dirty_segmap[type];
+ p->dirty_type = type;
p->ofs_unit = 1;
} else {
p->gc_mode = select_gc_type(gc_type);
p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
+ p->dirty_type = DIRTY;
p->ofs_unit = sbi->segs_per_sec;
}
p->offset = sbi->last_victim[p->gc_mode];
@@ -243,6 +245,8 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
struct victim_sel_policy p;
unsigned int secno, max_cost;
int nsearched = 0;
+ unsigned int max_search = MAX_VICTIM_SEARCH;
+ unsigned int nr_dirty;
p.alloc_mode = alloc_mode;
select_policy(sbi, gc_type, type, &p);
@@ -258,6 +262,27 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
goto got_it;
}
+ nr_dirty = dirty_i->nr_dirty[p.dirty_type];
+ if (p.gc_mode == GC_GREEDY && p.alloc_mode != SSR) {
+ if (TOTAL_SEGS(sbi) <= FULL_VICTIM_SEARCH_THRESH)
+ max_search = nr_dirty; /* search all the dirty segs */
+ else {
+ /*
+ * With more dirty segments, garbage blocks are likely
+ * more scattered, thus search harder for better
+ * victim.
+ */
+ max_search = div_u64 ((nr_dirty *
+ FULL_VICTIM_SEARCH_THRESH), TOTAL_SEGS(sbi));
+ if (max_search < MIN_VICTIM_SEARCH_GREEDY)
+ max_search = MIN_VICTIM_SEARCH_GREEDY;
+ }
+ }
+
+ /* no more than the total dirty segments */
+ if (max_search > nr_dirty)
+ max_search = nr_dirty;
+
while (1) {
unsigned long cost;
unsigned int segno;
@@ -290,7 +315,7 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
if (cost == max_cost)
continue;
- if (nsearched++ >= MAX_VICTIM_SEARCH) {
+ if (nsearched++ >= max_search) {
sbi->last_victim[p.gc_mode] = segno;
break;
}
diff --git a/fs/f2fs/gc.h b/fs/f2fs/gc.h
index 2c6a6bd..2f525aa 100644
--- a/fs/f2fs/gc.h
+++ b/fs/f2fs/gc.h
@@ -20,7 +20,9 @@
#define LIMIT_FREE_BLOCK 40 /* percentage over invalid + free space */
/* Search max. number of dirty segments to select a victim segment */
-#define MAX_VICTIM_SEARCH 20
+#define MAX_VICTIM_SEARCH 20
+#define MIN_VICTIM_SEARCH_GREEDY 20
+#define FULL_VICTIM_SEARCH_THRESH 4096
struct f2fs_gc_kthread {
struct task_struct *f2fs_gc_task;
diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
index 062424a..cd33f96 100644
--- a/fs/f2fs/segment.h
+++ b/fs/f2fs/segment.h
@@ -142,6 +142,7 @@ struct victim_sel_policy {
int alloc_mode; /* LFS or SSR */
int gc_mode; /* GC_CB or GC_GREEDY */
unsigned long *dirty_segmap; /* dirty segment bitmap */
+ int dirty_type;
unsigned int offset; /* last scanned bitmap offset */
unsigned int ofs_unit; /* bitmap search unit */
unsigned int min_cost; /* minimum cost */
--
1.7.9.5
Hi,
2013-08-29 (목), 08:48 +0800, Jin Xu:
> From: Jin Xu <[email protected]>
>
> This patch improves the foreground gc efficiency by optimizing the
> victim selection policy. With this optimization, the random re-write
> performance could increase up to 20%.
>
> For f2fs, foreground gc happens when disk is lack of free spaces,
> it selects dirty segments and moves valid blocks around for making
> more space available. The gc cost of a segment is determined by the
> valid blocks in the segment. The less the valid blocks, the higher
> the efficiency. The ideal victim segment is the one that has the
> most garbage blocks.
>
> Currently, it searches up to 20 dirty segments for a victim segment.
> The selected victim is not likely the best victim for gc when there
> are much more dirty segments. Why not searching more dirty segments
> for a better victim? The cost of searching dirty segments is
> negligible in comparison to moving blocks.
>
> In this patch, it does not search a constant number of dirty segments
> anymore, instead it calculates the number based on the total segments,
> dirty segments and a threshold. Following is the pseudo formula.
> ,-- nr_dirty_segments, if total_segments < threshold
> (# of search) = |
> `-- (nr_dirty_segments * threshold) / total_segments,
> Otherwise
Nice catch, but,
I don't understand why we search the # of segments in proportion to the
# of dirty segments.
How about the case where threshold = 10 and total_segments = 100000?
Or threshold = 1000000 and total_segments = 100?
For this, we need to define additional MIN/MAX thresholds and another
handling codes as your proposal.
>
> The test case is simple. It creates as many files until the disk full.
> The size for each file is 32KB. Then it writes as many as 100000
> records of 4KB size to random offsets of random files in sync mode.
> The testing was done on a 2GB partition of a SDHC card. Let's see the
> test result of f2fs without and with the patch.
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?
[snip]
> diff --git a/fs/f2fs/gc.c b/fs/f2fs/gc.c
> index 35f9b1a..4e045e6 100644
> --- a/fs/f2fs/gc.c
> +++ b/fs/f2fs/gc.c
> @@ -138,10 +138,12 @@ static void select_policy(struct f2fs_sb_info *sbi, int gc_type,
> if (p->alloc_mode == SSR) {
> p->gc_mode = GC_GREEDY;
> p->dirty_segmap = dirty_i->dirty_segmap[type];
> + p->dirty_type = type;
p->max_search = dirty_i->nr_dirty[type];
> p->ofs_unit = 1;
> } else {
> p->gc_mode = select_gc_type(gc_type);
> p->dirty_segmap = dirty_i->dirty_segmap[DIRTY];
> + p->dirty_type = DIRTY;
p->max_search = dirty_i->nr_dirty[DIRTY];
> p->ofs_unit = sbi->segs_per_sec;
> }
if (p->max_search > MAX_VICTIM_SEARCH)
p->max_search = MAX_VICTIM_SEARCH;
#define MAX_VICTIM_SEARCH 4096 /* covers 8GB */
> p->offset = sbi->last_victim[p->gc_mode];
> @@ -243,6 +245,8 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> struct victim_sel_policy p;
> unsigned int secno, max_cost;
> int nsearched = 0;
> + unsigned int max_search = MAX_VICTIM_SEARCH;
> + unsigned int nr_dirty;
>
> p.alloc_mode = alloc_mode;
> select_policy(sbi, gc_type, type, &p);
> @@ -258,6 +262,27 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> goto got_it;
> }
>
> + nr_dirty = dirty_i->nr_dirty[p.dirty_type];
> + if (p.gc_mode == GC_GREEDY && p.alloc_mode != SSR) {
> + if (TOTAL_SEGS(sbi) <= FULL_VICTIM_SEARCH_THRESH)
> + max_search = nr_dirty; /* search all the dirty segs */
> + else {
> + /*
> + * With more dirty segments, garbage blocks are likely
> + * more scattered, thus search harder for better
> + * victim.
> + */
> + max_search = div_u64 ((nr_dirty *
> + FULL_VICTIM_SEARCH_THRESH), TOTAL_SEGS(sbi));
> + if (max_search < MIN_VICTIM_SEARCH_GREEDY)
> + max_search = MIN_VICTIM_SEARCH_GREEDY;
> + }
> + }
> +
> + /* no more than the total dirty segments */
> + if (max_search > nr_dirty)
> + max_search = nr_dirty;
> +
> while (1) {
> unsigned long cost;
> unsigned int segno;
> @@ -290,7 +315,7 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> if (cost == max_cost)
> continue;
>
> - if (nsearched++ >= MAX_VICTIM_SEARCH) {
> + if (nsearched++ >= max_search) {
if (nsearched++ >= p.max_search) {
> sbi->last_victim[p.gc_mode] = segno;
> break;
> }
> diff --git a/fs/f2fs/gc.h b/fs/f2fs/gc.h
> index 2c6a6bd..2f525aa 100644
> --- a/fs/f2fs/gc.h
> +++ b/fs/f2fs/gc.h
> @@ -20,7 +20,9 @@
> #define LIMIT_FREE_BLOCK 40 /* percentage over invalid + free space */
>
> /* Search max. number of dirty segments to select a victim segment */
> -#define MAX_VICTIM_SEARCH 20
> +#define MAX_VICTIM_SEARCH 20
> +#define MIN_VICTIM_SEARCH_GREEDY 20
> +#define FULL_VICTIM_SEARCH_THRESH 4096
>
> struct f2fs_gc_kthread {
> struct task_struct *f2fs_gc_task;
> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
> index 062424a..cd33f96 100644
> --- a/fs/f2fs/segment.h
> +++ b/fs/f2fs/segment.h
> @@ -142,6 +142,7 @@ struct victim_sel_policy {
> int alloc_mode; /* LFS or SSR */
> int gc_mode; /* GC_CB or GC_GREEDY */
> unsigned long *dirty_segmap; /* dirty segment bitmap */
> + int dirty_type;
int max_search; /* maximum # of segments to search */
> unsigned int offset; /* last scanned bitmap offset */
> unsigned int ofs_unit; /* bitmap search unit */
> unsigned int min_cost; /* minimum cost */
--
Jaegeuk Kim
Samsung
Hi Jaegeuk,
On 08/29/2013 07:56 PM, Jaegeuk Kim wrote:
> Hi,
>
> 2013-08-29 (목), 08:48 +0800, Jin Xu:
>> From: Jin Xu <[email protected]>
>>
>> This patch improves the foreground gc efficiency by optimizing the
>> victim selection policy. With this optimization, the random re-write
>> performance could increase up to 20%.
>>
[...]
>>
>> In this patch, it does not search a constant number of dirty segments
>> anymore, instead it calculates the number based on the total segments,
>> dirty segments and a threshold. Following is the pseudo formula.
>> ,-- nr_dirty_segments, if total_segments < threshold
>> (# of search) = |
>> `-- (nr_dirty_segments * threshold) / total_segments,
>> Otherwise
>
> Nice catch, but,
> I don't understand why we search the # of segments in proportion to the
> # of dirty segments.
> How about the case where threshold = 10 and total_segments = 100000?
> Or threshold = 1000000 and total_segments = 100?
> For this, we need to define additional MIN/MAX thresholds and another
> handling codes as your proposal.
>
Firstly, calculating the # of search this way could constraint the
searching overhead in a certain level even when the total segments is
too many.
Secondly, when there are more dirty segments, the same # of garbage
blocks might be more scattered than when there are less dirty segments.
So we search the # of the segments in proportion to the # of dirty
segments.
[...]
>
> 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.
What do you think now?
> #define MAX_VICTIM_SEARCH 4096 /* covers 8GB */
>
>> p->offset = sbi->last_victim[p->gc_mode];
>> @@ -243,6 +245,8 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>> struct victim_sel_policy p;
>> unsigned int secno, max_cost;
>> int nsearched = 0;
>> + unsigned int max_search = MAX_VICTIM_SEARCH;
>> + unsigned int nr_dirty;
>>
>> p.alloc_mode = alloc_mode;
>> select_policy(sbi, gc_type, type, &p);
>> @@ -258,6 +262,27 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>> goto got_it;
>> }
>>
>> + nr_dirty = dirty_i->nr_dirty[p.dirty_type];
>> + if (p.gc_mode == GC_GREEDY && p.alloc_mode != SSR) {
>> + if (TOTAL_SEGS(sbi) <= FULL_VICTIM_SEARCH_THRESH)
>> + max_search = nr_dirty; /* search all the dirty segs */
>> + else {
>> + /*
>> + * With more dirty segments, garbage blocks are likely
>> + * more scattered, thus search harder for better
>> + * victim.
>> + */
>> + max_search = div_u64 ((nr_dirty *
>> + FULL_VICTIM_SEARCH_THRESH), TOTAL_SEGS(sbi));
>> + if (max_search < MIN_VICTIM_SEARCH_GREEDY)
>> + max_search = MIN_VICTIM_SEARCH_GREEDY;
>> + }
>> + }
>> +
>> + /* no more than the total dirty segments */
>> + if (max_search > nr_dirty)
>> + max_search = nr_dirty;
>> +
>> while (1) {
>> unsigned long cost;
>> unsigned int segno;
>> @@ -290,7 +315,7 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>> if (cost == max_cost)
>> continue;
>>
>> - if (nsearched++ >= MAX_VICTIM_SEARCH) {
>> + if (nsearched++ >= max_search) {
>
> if (nsearched++ >= p.max_search) {
>
>> sbi->last_victim[p.gc_mode] = segno;
>> break;
>> }
>> diff --git a/fs/f2fs/gc.h b/fs/f2fs/gc.h
>> index 2c6a6bd..2f525aa 100644
>> --- a/fs/f2fs/gc.h
>> +++ b/fs/f2fs/gc.h
>> @@ -20,7 +20,9 @@
>> #define LIMIT_FREE_BLOCK 40 /* percentage over invalid + free space */
>>
>> /* Search max. number of dirty segments to select a victim segment */
>> -#define MAX_VICTIM_SEARCH 20
>> +#define MAX_VICTIM_SEARCH 20
>> +#define MIN_VICTIM_SEARCH_GREEDY 20
>> +#define FULL_VICTIM_SEARCH_THRESH 4096
>>
>> struct f2fs_gc_kthread {
>> struct task_struct *f2fs_gc_task;
>> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
>> index 062424a..cd33f96 100644
>> --- a/fs/f2fs/segment.h
>> +++ b/fs/f2fs/segment.h
>> @@ -142,6 +142,7 @@ struct victim_sel_policy {
>> int alloc_mode; /* LFS or SSR */
>> int gc_mode; /* GC_CB or GC_GREEDY */
>> unsigned long *dirty_segmap; /* dirty segment bitmap */
>> + int dirty_type;
>
> int max_search; /* maximum # of segments to search */
>
>> unsigned int offset; /* last scanned bitmap offset */
>> unsigned int ofs_unit; /* bitmap search unit */
>> unsigned int min_cost; /* minimum cost */
>
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,
>
> What do you think now?
>
> > #define MAX_VICTIM_SEARCH 4096 /* covers 8GB */
> >
> >> p->offset = sbi->last_victim[p->gc_mode];
> >> @@ -243,6 +245,8 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> >> struct victim_sel_policy p;
> >> unsigned int secno, max_cost;
> >> int nsearched = 0;
> >> + unsigned int max_search = MAX_VICTIM_SEARCH;
> >> + unsigned int nr_dirty;
> >>
> >> p.alloc_mode = alloc_mode;
> >> select_policy(sbi, gc_type, type, &p);
> >> @@ -258,6 +262,27 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> >> goto got_it;
> >> }
> >>
> >> + nr_dirty = dirty_i->nr_dirty[p.dirty_type];
> >> + if (p.gc_mode == GC_GREEDY && p.alloc_mode != SSR) {
> >> + if (TOTAL_SEGS(sbi) <= FULL_VICTIM_SEARCH_THRESH)
> >> + max_search = nr_dirty; /* search all the dirty segs */
> >> + else {
> >> + /*
> >> + * With more dirty segments, garbage blocks are likely
> >> + * more scattered, thus search harder for better
> >> + * victim.
> >> + */
> >> + max_search = div_u64 ((nr_dirty *
> >> + FULL_VICTIM_SEARCH_THRESH), TOTAL_SEGS(sbi));
> >> + if (max_search < MIN_VICTIM_SEARCH_GREEDY)
> >> + max_search = MIN_VICTIM_SEARCH_GREEDY;
> >> + }
> >> + }
> >> +
> >> + /* no more than the total dirty segments */
> >> + if (max_search > nr_dirty)
> >> + max_search = nr_dirty;
> >> +
> >> while (1) {
> >> unsigned long cost;
> >> unsigned int segno;
> >> @@ -290,7 +315,7 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> >> if (cost == max_cost)
> >> continue;
> >>
> >> - if (nsearched++ >= MAX_VICTIM_SEARCH) {
> >> + if (nsearched++ >= max_search) {
> >
> > if (nsearched++ >= p.max_search) {
> >
> >> sbi->last_victim[p.gc_mode] = segno;
> >> break;
> >> }
> >> diff --git a/fs/f2fs/gc.h b/fs/f2fs/gc.h
> >> index 2c6a6bd..2f525aa 100644
> >> --- a/fs/f2fs/gc.h
> >> +++ b/fs/f2fs/gc.h
> >> @@ -20,7 +20,9 @@
> >> #define LIMIT_FREE_BLOCK 40 /* percentage over invalid + free space */
> >>
> >> /* Search max. number of dirty segments to select a victim segment */
> >> -#define MAX_VICTIM_SEARCH 20
> >> +#define MAX_VICTIM_SEARCH 20
> >> +#define MIN_VICTIM_SEARCH_GREEDY 20
> >> +#define FULL_VICTIM_SEARCH_THRESH 4096
> >>
> >> struct f2fs_gc_kthread {
> >> struct task_struct *f2fs_gc_task;
> >> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
> >> index 062424a..cd33f96 100644
> >> --- a/fs/f2fs/segment.h
> >> +++ b/fs/f2fs/segment.h
> >> @@ -142,6 +142,7 @@ struct victim_sel_policy {
> >> int alloc_mode; /* LFS or SSR */
> >> int gc_mode; /* GC_CB or GC_GREEDY */
> >> unsigned long *dirty_segmap; /* dirty segment bitmap */
> >> + int dirty_type;
> >
> > int max_search; /* maximum # of segments to search */
> >
> >> unsigned int offset; /* last scanned bitmap offset */
> >> unsigned int ofs_unit; /* bitmap search unit */
> >> unsigned int min_cost; /* minimum cost */
> >
>
--
Jaegeuk Kim
Samsung
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?
>>
>> What do you think now?
>>
>>> #define MAX_VICTIM_SEARCH 4096 /* covers 8GB */
>>>
>>>> p->offset = sbi->last_victim[p->gc_mode];
>>>> @@ -243,6 +245,8 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>>>> struct victim_sel_policy p;
>>>> unsigned int secno, max_cost;
>>>> int nsearched = 0;
>>>> + unsigned int max_search = MAX_VICTIM_SEARCH;
>>>> + unsigned int nr_dirty;
>>>>
>>>> p.alloc_mode = alloc_mode;
>>>> select_policy(sbi, gc_type, type, &p);
>>>> @@ -258,6 +262,27 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>>>> goto got_it;
>>>> }
>>>>
>>>> + nr_dirty = dirty_i->nr_dirty[p.dirty_type];
>>>> + if (p.gc_mode == GC_GREEDY && p.alloc_mode != SSR) {
>>>> + if (TOTAL_SEGS(sbi) <= FULL_VICTIM_SEARCH_THRESH)
>>>> + max_search = nr_dirty; /* search all the dirty segs */
>>>> + else {
>>>> + /*
>>>> + * With more dirty segments, garbage blocks are likely
>>>> + * more scattered, thus search harder for better
>>>> + * victim.
>>>> + */
>>>> + max_search = div_u64 ((nr_dirty *
>>>> + FULL_VICTIM_SEARCH_THRESH), TOTAL_SEGS(sbi));
>>>> + if (max_search < MIN_VICTIM_SEARCH_GREEDY)
>>>> + max_search = MIN_VICTIM_SEARCH_GREEDY;
>>>> + }
>>>> + }
>>>> +
>>>> + /* no more than the total dirty segments */
>>>> + if (max_search > nr_dirty)
>>>> + max_search = nr_dirty;
>>>> +
>>>> while (1) {
>>>> unsigned long cost;
>>>> unsigned int segno;
>>>> @@ -290,7 +315,7 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>>>> if (cost == max_cost)
>>>> continue;
>>>>
>>>> - if (nsearched++ >= MAX_VICTIM_SEARCH) {
>>>> + if (nsearched++ >= max_search) {
>>>
>>> if (nsearched++ >= p.max_search) {
>>>
>>>> sbi->last_victim[p.gc_mode] = segno;
>>>> break;
>>>> }
>>>> diff --git a/fs/f2fs/gc.h b/fs/f2fs/gc.h
>>>> index 2c6a6bd..2f525aa 100644
>>>> --- a/fs/f2fs/gc.h
>>>> +++ b/fs/f2fs/gc.h
>>>> @@ -20,7 +20,9 @@
>>>> #define LIMIT_FREE_BLOCK 40 /* percentage over invalid + free space */
>>>>
>>>> /* Search max. number of dirty segments to select a victim segment */
>>>> -#define MAX_VICTIM_SEARCH 20
>>>> +#define MAX_VICTIM_SEARCH 20
>>>> +#define MIN_VICTIM_SEARCH_GREEDY 20
>>>> +#define FULL_VICTIM_SEARCH_THRESH 4096
>>>>
>>>> struct f2fs_gc_kthread {
>>>> struct task_struct *f2fs_gc_task;
>>>> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
>>>> index 062424a..cd33f96 100644
>>>> --- a/fs/f2fs/segment.h
>>>> +++ b/fs/f2fs/segment.h
>>>> @@ -142,6 +142,7 @@ struct victim_sel_policy {
>>>> int alloc_mode; /* LFS or SSR */
>>>> int gc_mode; /* GC_CB or GC_GREEDY */
>>>> unsigned long *dirty_segmap; /* dirty segment bitmap */
>>>> + int dirty_type;
>>>
>>> int max_search; /* maximum # of segments to search */
>>>
>>>> unsigned int offset; /* last scanned bitmap offset */
>>>> unsigned int ofs_unit; /* bitmap search unit */
>>>> unsigned int min_cost; /* minimum cost */
>>>
>>
>
--
Thanks,
Jin
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?
Thanks,
>
> >>
> >> What do you think now?
> >>
> >>> #define MAX_VICTIM_SEARCH 4096 /* covers 8GB */
> >>>
> >>>> p->offset = sbi->last_victim[p->gc_mode];
> >>>> @@ -243,6 +245,8 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> >>>> struct victim_sel_policy p;
> >>>> unsigned int secno, max_cost;
> >>>> int nsearched = 0;
> >>>> + unsigned int max_search = MAX_VICTIM_SEARCH;
> >>>> + unsigned int nr_dirty;
> >>>>
> >>>> p.alloc_mode = alloc_mode;
> >>>> select_policy(sbi, gc_type, type, &p);
> >>>> @@ -258,6 +262,27 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> >>>> goto got_it;
> >>>> }
> >>>>
> >>>> + nr_dirty = dirty_i->nr_dirty[p.dirty_type];
> >>>> + if (p.gc_mode == GC_GREEDY && p.alloc_mode != SSR) {
> >>>> + if (TOTAL_SEGS(sbi) <= FULL_VICTIM_SEARCH_THRESH)
> >>>> + max_search = nr_dirty; /* search all the dirty segs */
> >>>> + else {
> >>>> + /*
> >>>> + * With more dirty segments, garbage blocks are likely
> >>>> + * more scattered, thus search harder for better
> >>>> + * victim.
> >>>> + */
> >>>> + max_search = div_u64 ((nr_dirty *
> >>>> + FULL_VICTIM_SEARCH_THRESH), TOTAL_SEGS(sbi));
> >>>> + if (max_search < MIN_VICTIM_SEARCH_GREEDY)
> >>>> + max_search = MIN_VICTIM_SEARCH_GREEDY;
> >>>> + }
> >>>> + }
> >>>> +
> >>>> + /* no more than the total dirty segments */
> >>>> + if (max_search > nr_dirty)
> >>>> + max_search = nr_dirty;
> >>>> +
> >>>> while (1) {
> >>>> unsigned long cost;
> >>>> unsigned int segno;
> >>>> @@ -290,7 +315,7 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
> >>>> if (cost == max_cost)
> >>>> continue;
> >>>>
> >>>> - if (nsearched++ >= MAX_VICTIM_SEARCH) {
> >>>> + if (nsearched++ >= max_search) {
> >>>
> >>> if (nsearched++ >= p.max_search) {
> >>>
> >>>> sbi->last_victim[p.gc_mode] = segno;
> >>>> break;
> >>>> }
> >>>> diff --git a/fs/f2fs/gc.h b/fs/f2fs/gc.h
> >>>> index 2c6a6bd..2f525aa 100644
> >>>> --- a/fs/f2fs/gc.h
> >>>> +++ b/fs/f2fs/gc.h
> >>>> @@ -20,7 +20,9 @@
> >>>> #define LIMIT_FREE_BLOCK 40 /* percentage over invalid + free space */
> >>>>
> >>>> /* Search max. number of dirty segments to select a victim segment */
> >>>> -#define MAX_VICTIM_SEARCH 20
> >>>> +#define MAX_VICTIM_SEARCH 20
> >>>> +#define MIN_VICTIM_SEARCH_GREEDY 20
> >>>> +#define FULL_VICTIM_SEARCH_THRESH 4096
> >>>>
> >>>> struct f2fs_gc_kthread {
> >>>> struct task_struct *f2fs_gc_task;
> >>>> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
> >>>> index 062424a..cd33f96 100644
> >>>> --- a/fs/f2fs/segment.h
> >>>> +++ b/fs/f2fs/segment.h
> >>>> @@ -142,6 +142,7 @@ struct victim_sel_policy {
> >>>> int alloc_mode; /* LFS or SSR */
> >>>> int gc_mode; /* GC_CB or GC_GREEDY */
> >>>> unsigned long *dirty_segmap; /* dirty segment bitmap */
> >>>> + int dirty_type;
> >>>
> >>> int max_search; /* maximum # of segments to search */
> >>>
> >>>> unsigned int offset; /* last scanned bitmap offset */
> >>>> unsigned int ofs_unit; /* bitmap search unit */
> >>>> unsigned int min_cost; /* minimum cost */
> >>>
> >>
> >
>
> --
> Thanks,
> Jin
> --
> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
--
Jaegeuk Kim
Samsung
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?
--
Thanks,
Jin
>
> Thanks,
>
>>
>>>>
>>>> What do you think now?
>>>>
>>>>> #define MAX_VICTIM_SEARCH 4096 /* covers 8GB */
>>>>>
>>>>>> p->offset = sbi->last_victim[p->gc_mode];
>>>>>> @@ -243,6 +245,8 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>>>>>> struct victim_sel_policy p;
>>>>>> unsigned int secno, max_cost;
>>>>>> int nsearched = 0;
>>>>>> + unsigned int max_search = MAX_VICTIM_SEARCH;
>>>>>> + unsigned int nr_dirty;
>>>>>>
>>>>>> p.alloc_mode = alloc_mode;
>>>>>> select_policy(sbi, gc_type, type, &p);
>>>>>> @@ -258,6 +262,27 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>>>>>> goto got_it;
>>>>>> }
>>>>>>
>>>>>> + nr_dirty = dirty_i->nr_dirty[p.dirty_type];
>>>>>> + if (p.gc_mode == GC_GREEDY && p.alloc_mode != SSR) {
>>>>>> + if (TOTAL_SEGS(sbi) <= FULL_VICTIM_SEARCH_THRESH)
>>>>>> + max_search = nr_dirty; /* search all the dirty segs */
>>>>>> + else {
>>>>>> + /*
>>>>>> + * With more dirty segments, garbage blocks are likely
>>>>>> + * more scattered, thus search harder for better
>>>>>> + * victim.
>>>>>> + */
>>>>>> + max_search = div_u64 ((nr_dirty *
>>>>>> + FULL_VICTIM_SEARCH_THRESH), TOTAL_SEGS(sbi));
>>>>>> + if (max_search < MIN_VICTIM_SEARCH_GREEDY)
>>>>>> + max_search = MIN_VICTIM_SEARCH_GREEDY;
>>>>>> + }
>>>>>> + }
>>>>>> +
>>>>>> + /* no more than the total dirty segments */
>>>>>> + if (max_search > nr_dirty)
>>>>>> + max_search = nr_dirty;
>>>>>> +
>>>>>> while (1) {
>>>>>> unsigned long cost;
>>>>>> unsigned int segno;
>>>>>> @@ -290,7 +315,7 @@ static int get_victim_by_default(struct f2fs_sb_info *sbi,
>>>>>> if (cost == max_cost)
>>>>>> continue;
>>>>>>
>>>>>> - if (nsearched++ >= MAX_VICTIM_SEARCH) {
>>>>>> + if (nsearched++ >= max_search) {
>>>>>
>>>>> if (nsearched++ >= p.max_search) {
>>>>>
>>>>>> sbi->last_victim[p.gc_mode] = segno;
>>>>>> break;
>>>>>> }
>>>>>> diff --git a/fs/f2fs/gc.h b/fs/f2fs/gc.h
>>>>>> index 2c6a6bd..2f525aa 100644
>>>>>> --- a/fs/f2fs/gc.h
>>>>>> +++ b/fs/f2fs/gc.h
>>>>>> @@ -20,7 +20,9 @@
>>>>>> #define LIMIT_FREE_BLOCK 40 /* percentage over invalid + free space */
>>>>>>
>>>>>> /* Search max. number of dirty segments to select a victim segment */
>>>>>> -#define MAX_VICTIM_SEARCH 20
>>>>>> +#define MAX_VICTIM_SEARCH 20
>>>>>> +#define MIN_VICTIM_SEARCH_GREEDY 20
>>>>>> +#define FULL_VICTIM_SEARCH_THRESH 4096
>>>>>>
>>>>>> struct f2fs_gc_kthread {
>>>>>> struct task_struct *f2fs_gc_task;
>>>>>> diff --git a/fs/f2fs/segment.h b/fs/f2fs/segment.h
>>>>>> index 062424a..cd33f96 100644
>>>>>> --- a/fs/f2fs/segment.h
>>>>>> +++ b/fs/f2fs/segment.h
>>>>>> @@ -142,6 +142,7 @@ struct victim_sel_policy {
>>>>>> int alloc_mode; /* LFS or SSR */
>>>>>> int gc_mode; /* GC_CB or GC_GREEDY */
>>>>>> unsigned long *dirty_segmap; /* dirty segment bitmap */
>>>>>> + int dirty_type;
>>>>>
>>>>> int max_search; /* maximum # of segments to search */
>>>>>
>>>>>> unsigned int offset; /* last scanned bitmap offset */
>>>>>> unsigned int ofs_unit; /* bitmap search unit */
>>>>>> unsigned int min_cost; /* minimum cost */
>>>>>
>>>>
>>>
>>
>> --
>> Thanks,
>> Jin
>> --
>> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
>> the body of a message to [email protected]
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
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,
--
Jaegeuk Kim
Samsung
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,
>