On Mon, Feb 01, 2021 at 11:52:48AM +0900, Changheun Lee wrote:
> > On Fri, Jan 29, 2021 at 12:49:08PM +0900, Changheun Lee wrote:
> > > bio size can grow up to 4GB when muli-page bvec is enabled.
> > > but sometimes it would lead to inefficient behaviors.
> > > in case of large chunk direct I/O, - 32MB chunk read in user space -
> > > all pages for 32MB would be merged to a bio structure if the pages
> > > physical addresses are contiguous. it makes some delay to submit
> > > until merge complete. bio max size should be limited to a proper size.
> > >
> > > When 32MB chunk read with direct I/O option is coming from userspace,
> > > kernel behavior is below now. it's timeline.
> > >
> > > | bio merge for 32MB. total 8,192 pages are merged.
> > > | total elapsed time is over 2ms.
> > > |------------------ ... ----------------------->|
> > > | 8,192 pages merged a bio.
> > > | at this time, first bio submit is done.
> > > | 1 bio is split to 32 read request and issue.
> > > |--------------->
> > > |--------------->
> > > |--------------->
> > > ......
> > > |--------------->
> > > |--------------->|
> > > total 19ms elapsed to complete 32MB read done from device. |
> > >
> > > If bio max size is limited with 1MB, behavior is changed below.
> > >
bio_iov_iter_get_pages> > > | bio merge for 1MB. 256 pages are merged for each bio.
> > > | total 32 bio will be made.
> > > | total elapsed time is over 2ms. it's same.
> > > | but, first bio submit timing is fast. about 100us.
> > > |--->|--->|--->|---> ... -->|--->|--->|--->|--->|
> > > | 256 pages merged a bio.
> > > | at this time, first bio submit is done.
> > > | and 1 read request is issued for 1 bio.
> > > |--------------->
> > > |--------------->
> > > |--------------->
> > > ......
> > > |--------------->
> > > |--------------->|
> > > total 17ms elapsed to complete 32MB read done from device. |
> >
> > Can you share us if enabling THP in your application can avoid this issue? BTW, you
> > need to make the 32MB buffer aligned with huge page size. IMO, THP perfectly fits
> > your case.
> >
>
> THP is enabled already like as below in my environment. It has no effect.
>
> cat /sys/kernel/mm/transparent_hugepage/enabled
> [always] madvise never
The 32MB user buffer needs to be huge page size aligned. If your system
supports bcc/bpftrace, it is quite easy to check if the buffer is
aligned.
>
> This issue was reported from performance benchmark application in open market.
> I can't control application's working in open market.
> It's not only my own case. This issue might be occured in many mobile environment.
> At least, I checked this problem in exynos, and qualcomm chipset.
You just said it takes 2ms for building 32MB bio, but you never investigate the
reason. I guess it is from get_user_pages_fast(), but maybe others. Can you dig
further for the reason? Maybe it is one arm64 specific issue.
BTW, bio_iov_iter_get_pages() just takes ~200us on one x86_64 VM with THP, which is
observed via bcc/funclatency when running the following workload:
[root@ktest-01 test]# cat fio.job
[global]
bs=32768k
rw=randread
iodepth=1
ioengine=psync
direct=1
runtime=20
time_based
group_reporting=0
ramp_time=5
[diotest]
filename=/dev/sde
[root@ktest-01 func]# /usr/share/bcc/tools/funclatency bio_iov_iter_get_pages
Tracing 1 functions for "bio_iov_iter_get_pages"... Hit Ctrl-C to end.
^C
nsecs : count distribution
0 -> 1 : 0 | |
2 -> 3 : 0 | |
4 -> 7 : 0 | |
8 -> 15 : 0 | |
16 -> 31 : 0 | |
32 -> 63 : 0 | |
64 -> 127 : 0 | |
128 -> 255 : 0 | |
256 -> 511 : 0 | |
512 -> 1023 : 0 | |
1024 -> 2047 : 0 | |
2048 -> 4095 : 0 | |
4096 -> 8191 : 0 | |
8192 -> 16383 : 0 | |
16384 -> 32767 : 0 | |
32768 -> 65535 : 0 | |
65536 -> 131071 : 0 | |
131072 -> 262143 : 1842 |****************************************|
262144 -> 524287 : 125 |** |
524288 -> 1048575 : 6 | |
1048576 -> 2097151 : 0 | |
2097152 -> 4194303 : 1 | |
4194304 -> 8388607 : 0 | |
8388608 -> 16777215 : 1 | |
Detaching...
--
Ming
