RE: [f2fs-dev] [PATCH v2] f2fs: serialize block allocation of dio writes to enhance multithread performance

From: Chao Yu
Date: Fri Jan 15 2016 - 03:52:32 EST

Next message: Ludovic Desroches: "[RESEND PATCH v3 2/5] MAINTAINERS: add entry for Atmel SAMA5D2 ADC driver"
Previous message: Peter Zijlstra: "Re: [RFC 8/8] Do not reclaim the whole CPU bandwidth"
In reply to: Jaegeuk Kim: "Re: [f2fs-dev] [PATCH v2] f2fs: serialize block allocation of dio writes to enhance multithread performance"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

Hi Jaegeuk,

> -----Original Message-----
> From: Jaegeuk Kim [mailto:jaegeuk@xxxxxxxxxx]
> Sent: Friday, January 15, 2016 7:54 AM
> To: Chao Yu
> Cc: linux-kernel@xxxxxxxxxxxxxxx; linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx
> Subject: Re: [f2fs-dev] [PATCH v2] f2fs: serialize block allocation of dio writes to enhance
> multithread performance
>
> Hi Chao,
>
> On Wed, Jan 13, 2016 at 08:13:36PM +0800, Chao Yu wrote:
> > Hi Jaegeuk,
> >
> > > -----Original Message-----
> > > From: Chao Yu [mailto:chao2.yu@xxxxxxxxxxx]
> > > Sent: Tuesday, December 29, 2015 2:44 PM
> > > To: Jaegeuk Kim
> > > Cc: linux-kernel@xxxxxxxxxxxxxxx; linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx
> > > Subject: [f2fs-dev] [PATCH v2] f2fs: serialize block allocation of dio writes to enhance
> > > multithread performance
> > >
> > > When performing big dio writes concurrently, our performace will be low
> > > because of Thread A's allocation of multi continuous blocks will be
> > > interrupted by Thread B, there are two cases as below:
> > > - In Thread B, we may change current segment to a new segment for LFS
> > > allocation if we dio write in the beginning of the file.
> > > - In Thread B, we may allocate blocks in the middle of Thread A's
> > > allocation, which make blocks allocated in Thread A being inconsecutive.
> > >
> > > This patch adds writepages mutex lock to make block allocation in dio write
> > > being atomic to avoid above issues.
> > >
> > > Test environment 1:
> > > ubuntu os with linux kernel 4.4-rc4, intel i7-3770, 16g memory,
> > > 32g kingston sd card.
> > >
> > > fio --name seqw --ioengine=sync --invalidate=1 --rw=write --directory=/mnt/f2fs
> > > --filesize=256m --size=16m --bs=2m --direct=1
> > > --numjobs=10
> > >
> > > before:
> > > WRITE: io=163840KB, aggrb=5125KB/s, minb=512KB/s, maxb=776KB/s, mint=21105msec,
> > > maxt=31967msec
> > > patched:
> > > WRITE: io=163840KB, aggrb=10424KB/s, minb=1042KB/s, maxb=1172KB/s, mint=13975msec,
> > > maxt=15717msec
> > >
> > > Test environment 2:
> > > Note4 eMMC
> > >
> > > fio --name seqw --ioengine=sync --invalidate=1 --rw=write --directory=/data/test/
> > > --filesize=256m --size=64m --bs=2m --direct=1
> > > --numjobs=16
> > >
> > > before:
> > > WRITE: io=1024.0MB, aggrb=103583KB/s, minb=6473KB/s, maxb=8806KB/s, mint=7442msec,
> > > maxt=10123msec
> > > patched:
> > > WRITE: io=1024.0MB, aggrb=124860KB/s, minb=7803KB/s, maxb=9315KB/s, mint=7035msec,
> > > maxt=8398msec
> > >
> > > As Yunlei He reported when he test with current patch:
> > > "Does share writepages mutex lock have an effect on cache write?
> > > Here is AndroBench result on my phone:
> > >
> > > Before patch:
> > > 1R1W 8R8W 16R16W
> > > Sequential Write 161.31 163.85 154.67
> > > Random Write 9.48 17.66 18.09
> > >
> > > After patch:
> > > 1R1W 8R8W 16R16W
> > > Sequential Write 159.61 157.24 160.11
> > > Random Write 9.17 8.51 8.8
> > >
> > > Unit:Mb/s, File size: 64M, Buffer size: 4k"
> > >
> > > The turth is androidbench uses single thread with dio write to test performance
> > > of sequential write, and use multi-threads with dio write to test performance
> > > of random write. so we can not see any improvement in sequentail write test
> > > since serializing dio page allocation can only improve performance in
> > > multi-thread scenario, and there is a regression in multi-thread test with 4k
> > > dio write, this is because grabbing sbi->writepages lock for serializing block
> > > allocation stop the concurrency, so that less small dio bios could be merged,
> > > moreover, when there are huge number of small dio writes, grabbing mutex lock
> > > per dio increases the overhead.
> >
> > Since the whole DIOs in Androbench are IPU, so in mutex_lock/allocate_block/
> > mutex_unlock actually we didn't allocating blocks, but blocking the concurrency
> > by grabing sbi->writepages.
>
> Hmm, DIO doesn't grab writepages.

Yep, writepages mutex lock has not been added in ->direct_IO yet.

Previously, I'm intended to add this lock into ->direct_IO, but as Yunlei
reported, there would be a regression when he tested that patch with
Androbench.

So, what I said above is try to explain the reason why it caused a
regression.

Below is test result which I test current v2 patch ("f2fs: serialize block
allocation of dio writes to enhance multithread performance") in Note4.

a) DIO Append Write:
IO size: 4K
fio --name seqw --ioengine=sync --invalidate=1 --rw=write --directory=/data/test/ --filesize=16m --size=16m --bs=4k --direct=1
--numjobs=32
serialize all serialize none
threshold 1 2
1 45637 42753
2 40101 39263
3 44423 34576
4 43426 38701
5 39730 39000
6 39927 40870
7 41845 38099
8 41936 37691
average 42128.125 38869.125

IO size: 128K
fio --name seqw --ioengine=sync --invalidate=1 --rw=write --directory=/data/test/ --filesize=32m --size=32m --bs=128k --direct=1
--numjobs=32
serialize all serialize none
threshold 32 33
1 133712 105660
2 134899 113765
3 134969 100669
4 134157 117711
5 134657 114988
6 134277 112871
7 134761 116651
8 133270 118764
average 134337.75 112634.875

IO size: 2M
fio --name seqw --ioengine=sync --invalidate=1 --rw=write --directory=/data/test/ --filesize=64m --size=64m --bs=2m --direct=1
--numjobs=32
serialize all serialize none
threshold 512 513
1 133542 128572
2 133934 130843
3 132714 130566
4 132437 130663
5 133068 130826
6 132362 132120
7 134149 129645
8 132975 130080
average 133147.625 130414.375

IO size: 4K-1024K
fio --name seqw --ioengine=sync --invalidate=1 --rw=write --directory=/data/test/ --filesize=32m --size=32m --bsrang=4k-1024k
--direct=1 --numjobs=32
serialize all serialize partial serialize none
threshold 1 33 257
1 118724 101922 109226
2 114174 108503 108435
3 115673 109890 106605
4 111824 112762 114648
5 110144 102300 115253
6 108626 107063 113963
7 114824 113469 108234
8 102610 110609 109397
average 112074.875 108314.75 110720.125

b) DIO Random Overwrite:
IO size: 4K
fio --name randw --ioengine=sync --invalidate=1 --rw=randwrite --directory=/data/test/ --filesize=64m --size=16m --bs=4k --direct=1
--numjobs=32
serialize all serialize none
threshold 1 2
1 17454 17340
2 16266 17829
3 17167 17286
4 16746 17578
5 17250 18148
6 17260 17589
7 17146 16568
8 17106 17725
average 17049.375 17507.875

IO size: 16K
fio --name randw --ioengine=sync --invalidate=1 --rw=randwrite --directory=/data/test/ --filesize=64m --size=16m --bs=32k --direct=1
--numjobs=32
serialize all serialize none
threshold 8 9
1 47779 39845
2 34887 42638
3 46987 42380
4 45138 46500
5 41862 49044
6 40061 49643
7 42729 43647
8 40441 46030
average 42485.5 44965.875

IO size: 4K-1024K
fio --name randw --ioengine=sync --invalidate=1 --rw=randwrite --directory=/data/test/ --filesize=64m --size=16m --bsrang=4k-1024k
--direct=1 --numjobs=32
serialize all serialize partial serialize none
threshold 1 33 257
1 91235 95528 80236
2 90882 86887 94293
3 83446 92117 89710
4 89083 84344 92902
5 94335 85906 94487
6 84371 95779 89399
7 95597 87964 89264
8 94352 94394 92374
average 90412.625 90364.875 90333.125

We can see in all cases of append DIO write, serialization of DIOs will
lead to better performance, so how do you think of adding writepages lock
when writing offset is larger than isize?

> And, if we need to check every time, we can lose the bandwidth on every new
> appending writes.
>
> How about adding a condition to check whether we need to allocate or not?
>
> For example,
> if (i_size == calculated i_blocks && offset < i_size)
> skip allocation;

Yeah, that could be a faster way, I will take this. And I expect 'fallocate
after isize' case or 'punch inside isize' case will be rare one, then our
judgment condition will be accurate most of the time. :)

Thanks,

>
> Thanks,
>
> >
> > How about grabbing sbi->writepages when it needs to allocate block(s) like
> > following patch?
> >
> > ---
> > fs/f2fs/data.c | 58 +++++++++++++++++++++++++++++++++++++++++++++++++++-------
> > 1 file changed, 51 insertions(+), 7 deletions(-)
> >
> > diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> > index ac9e7c6..4439a85 100644
> > --- a/fs/f2fs/data.c
> > +++ b/fs/f2fs/data.c
> > @@ -458,6 +458,44 @@ got_it:
> > return page;
> > }
> >
> > +static bool __need_allocate_blocks(struct inode *inode, loff_t offset,
> > + size_t count)
> > +{
> > + struct dnode_of_data dn;
> > + pgoff_t start = F2FS_BYTES_TO_BLK(offset);
> > + pgoff_t end = start + F2FS_BYTES_TO_BLK(count);
> > + pgoff_t end_offset;
> > +
> > + while (start < end) {
> > + int ret;
> > +
> > + set_new_dnode(&dn, inode, NULL, NULL, 0);
> > + ret = get_dnode_of_data(&dn, start, LOOKUP_NODE);
> > + if (ret == -ENOENT) {
> > + return true;
> > + } else if (ret) {
> > + return ret;
> > + }
> > +
> > + end_offset = ADDRS_PER_PAGE(dn.node_page, F2FS_I(inode));
> > +
> > + while (dn.ofs_in_node < end_offset && start < end) {
> > + block_t blkaddr;
> > +
> > + blkaddr = datablock_addr(dn.node_page, dn.ofs_in_node);
> > + if (blkaddr == NULL_ADDR || blkaddr == NEW_ADDR) {
> > + f2fs_put_dnode(&dn);
> > + return true;
> > + }
> > + start++;
> > + dn.ofs_in_node++;
> > + }
> > +
> > + f2fs_put_dnode(&dn);
> > + }
> > + return false
> > +}
> > +
> > static int __allocate_data_block(struct dnode_of_data *dn)
> > {
> > struct f2fs_sb_info *sbi = F2FS_I_SB(dn->inode);
> > @@ -1620,7 +1658,9 @@ static ssize_t f2fs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
> > struct file *file = iocb->ki_filp;
> > struct address_space *mapping = file->f_mapping;
> > struct inode *inode = mapping->host;
> > + struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> > size_t count = iov_iter_count(iter);
> > + int rw = iov_iter_rw(iter);
> > int err;
> >
> > /* we don't need to use inline_data strictly */
> > @@ -1635,20 +1675,24 @@ static ssize_t f2fs_direct_IO(struct kiocb *iocb, struct iov_iter
> *iter,
> > if (err)
> > return err;
> >
> > - trace_f2fs_direct_IO_enter(inode, offset, count, iov_iter_rw(iter));
> > + trace_f2fs_direct_IO_enter(inode, offset, count, rw);
> >
> > - if (iov_iter_rw(iter) == WRITE) {
> > - err = __allocate_data_blocks(inode, offset, count);
> > - if (err)
> > - goto out;
> > + if (rw == WRITE) {
> > + if (__need_allocate_blocks(inode, offset, count) > 0) {
> > + mutex_lock(&sbi->writepages);
> > + err = __allocate_data_blocks(inode, offset, count);
> > + mutex_unlock(&sbi->writepages);
> > + if (err)
> > + goto out;
> > + }
> > }
> >
> > err = blockdev_direct_IO(iocb, inode, iter, offset, get_data_block_dio);
> > out:
> > - if (err < 0 && iov_iter_rw(iter) == WRITE)
> > + if (err < 0 && rw == WRITE)
> > f2fs_write_failed(mapping, offset + count);
> >
> > - trace_f2fs_direct_IO_exit(inode, offset, count, iov_iter_rw(iter), err);
> > + trace_f2fs_direct_IO_exit(inode, offset, count, rw, err);
> >
> > return err;
> > }
> > --
> > 2.6.3
> >
> > Thanks,
> >
> > >
> > > After all, serializing dio could only be used for concurrent scenario of
> > > big dio, so this patch also introduces a threshold in sysfs to provide user
> > > the interface of defining 'a big dio' with specified page number, which could
> > > be used to control wthether serialize or not that kind of dio with specified
> > > page number.
> > >
> > > The optimization works in rare scenario.
> > >
> > > Signed-off-by: Chao Yu <chao2.yu@xxxxxxxxxxx>
> > > ---
> > > v2:
> > > - merge another related patch into this one.
> > > ---
> > > Documentation/ABI/testing/sysfs-fs-f2fs | 12 ++++++++++++
> > > fs/f2fs/data.c | 17 +++++++++++++----
> > > fs/f2fs/f2fs.h | 3 +++
> > > fs/f2fs/super.c | 3 +++
> > > 4 files changed, 31 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/Documentation/ABI/testing/sysfs-fs-f2fs
> > > b/Documentation/ABI/testing/sysfs-fs-f2fs
> > > index 0345f2d..560a4f1 100644
> > > --- a/Documentation/ABI/testing/sysfs-fs-f2fs
> > > +++ b/Documentation/ABI/testing/sysfs-fs-f2fs
> > > @@ -92,3 +92,15 @@ Date: October 2015
> > > Contact: "Chao Yu" <chao2.yu@xxxxxxxxxxx>
> > > Description:
> > > Controls the count of nid pages to be readaheaded.
> > > +
> > > +What: /sys/fs/f2fs/<disk>/serialized_dio_pages
> > > +Date: December 2015
> > > +Contact: "Chao Yu" <chao2.yu@xxxxxxxxxxx>
> > > +Description:
> > > + It is a threshold with the unit of page size.
> > > + If DIO page count is equal or big than the threshold,
> > > + whole process of block address allocation of dio pages
> > > + will become atomic like buffered write.
> > > + It is used to maximize bandwidth utilization in the
> > > + scenario of concurrent write with dio vs buffered or
> > > + dio vs dio.
> > > diff --git a/fs/f2fs/data.c b/fs/f2fs/data.c
> > > index 8a89810..d506a0e 100644
> > > --- a/fs/f2fs/data.c
> > > +++ b/fs/f2fs/data.c
> > > @@ -1619,7 +1619,9 @@ static ssize_t f2fs_direct_IO(struct kiocb *iocb, struct iov_iter
> *iter,
> > > struct file *file = iocb->ki_filp;
> > > struct address_space *mapping = file->f_mapping;
> > > struct inode *inode = mapping->host;
> > > + struct f2fs_sb_info *sbi = F2FS_I_SB(inode);
> > > size_t count = iov_iter_count(iter);
> > > + int rw = iov_iter_rw(iter);
> > > int err;
> > >
> > > /* we don't need to use inline_data strictly */
> > > @@ -1634,20 +1636,27 @@ static ssize_t f2fs_direct_IO(struct kiocb *iocb, struct iov_iter
> *iter,
> > > if (err)
> > > return err;
> > >
> > > - trace_f2fs_direct_IO_enter(inode, offset, count, iov_iter_rw(iter));
> > > + trace_f2fs_direct_IO_enter(inode, offset, count, rw);
> > > +
> > > + if (rw == WRITE) {
> > > + bool serialized = (F2FS_BYTES_TO_BLK(count) >=
> > > + sbi->serialized_dio_pages);
> > >
> > > - if (iov_iter_rw(iter) == WRITE) {
> > > + if (serialized)
> > > + mutex_lock(&sbi->writepages);
> > > err = __allocate_data_blocks(inode, offset, count);
> > > + if (serialized)
> > > + mutex_unlock(&sbi->writepages);
> > > if (err)
> > > goto out;
> > > }
> > >
> > > err = blockdev_direct_IO(iocb, inode, iter, offset, get_data_block_dio);
> > > out:
> > > - if (err < 0 && iov_iter_rw(iter) == WRITE)
> > > + if (err < 0 && rw == WRITE)
> > > f2fs_write_failed(mapping, offset + count);
> > >
> > > - trace_f2fs_direct_IO_exit(inode, offset, count, iov_iter_rw(iter), err);
> > > + trace_f2fs_direct_IO_exit(inode, offset, count, rw, err);
> > >
> > > return err;
> > > }
> > > diff --git a/fs/f2fs/f2fs.h b/fs/f2fs/f2fs.h
> > > index a339508..293dc4e 100644
> > > --- a/fs/f2fs/f2fs.h
> > > +++ b/fs/f2fs/f2fs.h
> > > @@ -333,6 +333,8 @@ enum {
> > >
> > > #define MAX_DIR_RA_PAGES 4 /* maximum ra pages of dir */
> > >
> > > +#define DEF_SERIALIZED_DIO_PAGES 64 /* default serialized dio pages */
> > > +
> > > /* vector size for gang look-up from extent cache that consists of radix tree */
> > > #define EXT_TREE_VEC_SIZE 64
> > >
> > > @@ -784,6 +786,7 @@ struct f2fs_sb_info {
> > > unsigned int total_valid_inode_count; /* valid inode count */
> > > int active_logs; /* # of active logs */
> > > int dir_level; /* directory level */
> > > + int serialized_dio_pages; /* serialized direct IO pages */
> > >
> > > block_t user_block_count; /* # of user blocks */
> > > block_t total_valid_block_count; /* # of valid blocks */
> > > diff --git a/fs/f2fs/super.c b/fs/f2fs/super.c
> > > index a2e3a8f..4a2e51e 100644
> > > --- a/fs/f2fs/super.c
> > > +++ b/fs/f2fs/super.c
> > > @@ -218,6 +218,7 @@ F2FS_RW_ATTR(NM_INFO, f2fs_nm_info, ram_thresh, ram_thresh);
> > > F2FS_RW_ATTR(NM_INFO, f2fs_nm_info, ra_nid_pages, ra_nid_pages);
> > > F2FS_RW_ATTR(F2FS_SBI, f2fs_sb_info, max_victim_search, max_victim_search);
> > > F2FS_RW_ATTR(F2FS_SBI, f2fs_sb_info, dir_level, dir_level);
> > > +F2FS_RW_ATTR(F2FS_SBI, f2fs_sb_info, serialized_dio_pages, serialized_dio_pages);
> > > F2FS_RW_ATTR(F2FS_SBI, f2fs_sb_info, cp_interval, cp_interval);
> > >
> > > #define ATTR_LIST(name) (&f2fs_attr_##name.attr)
> > > @@ -234,6 +235,7 @@ static struct attribute *f2fs_attrs[] = {
> > > ATTR_LIST(min_fsync_blocks),
> > > ATTR_LIST(max_victim_search),
> > > ATTR_LIST(dir_level),
> > > + ATTR_LIST(serialized_dio_pages),
> > > ATTR_LIST(ram_thresh),
> > > ATTR_LIST(ra_nid_pages),
> > > ATTR_LIST(cp_interval),
> > > @@ -1125,6 +1127,7 @@ static void init_sb_info(struct f2fs_sb_info *sbi)
> > > atomic_set(&sbi->nr_pages[i], 0);
> > >
> > > sbi->dir_level = DEF_DIR_LEVEL;
> > > + sbi->serialized_dio_pages = DEF_SERIALIZED_DIO_PAGES;
> > > sbi->cp_interval = DEF_CP_INTERVAL;
> > > clear_sbi_flag(sbi, SBI_NEED_FSCK);
> > >
> > > --
> > > 2.6.3
> > >
> > >
> > >
> > > ------------------------------------------------------------------------------
> > > _______________________________________________
> > > Linux-f2fs-devel mailing list
> > > Linux-f2fs-devel@xxxxxxxxxxxxxxxxxxxxx
> > > https://lists.sourceforge.net/lists/listinfo/linux-f2fs-devel

Next message: Ludovic Desroches: "[RESEND PATCH v3 2/5] MAINTAINERS: add entry for Atmel SAMA5D2 ADC driver"
Previous message: Peter Zijlstra: "Re: [RFC 8/8] Do not reclaim the whole CPU bandwidth"
In reply to: Jaegeuk Kim: "Re: [f2fs-dev] [PATCH v2] f2fs: serialize block allocation of dio writes to enhance multithread performance"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]