On 01/17/2018 10:29 PM, Mike Snitzer wrote:
On Wed, Jan 17 2018 at 2:33pm -0500,
Heinz Mauelshagen <heinzm@xxxxxxxxxx> wrote:
Enhancing IO performance compared to the kernels existing
loop driver thus better suiting respective requirements in
test setups, this patch series adds new "loop" and "ram" targets.
For measures see test results below.
The "loop" target maps segments to backing files.
Mapping table example:
0 4192256 loop /tmp/mapper_loop1
4192256 2097152 loop /dev/nvm/mapper_loop0
The "ram" target accesses RAM directly rather than through
tmpfs additionally enhancing performance compared to "loop"
thus avoding filesystem overhead.
Mapping table example:
0 8388608 ram
"ram" is a singleton target.
Performance test results for 4K and 32K IOPS comparing the loop driver
with dm-loop backed by tmpfs and dm-ram (all 2GiB backing size):
<TESTSCRIPT>
#!/bin/sh
for f in /tmp/loop0 /tmp/mapper_loop0
do
dd if=/dev/zero of=$f bs=256M count=8 iflag=fullblock
done
losetup /dev/loop0 /tmp/loop0
sectors=`du -s /tmp/mapper_loop0|cut -f1`
dmsetup create loop0 --table "0 $sectors loop /tmp/mapper_loop0"
dmsetup create ram --table "0 $sectors ram"
for bs in 4K 32K
do
for d in /dev/loop0 /dev/mapper/loop0 /dev/mapper/ram
do
echo 3 > /proc/sys/vm/drop_caches
fio --bs=$bs --rw=randrw --numjobs=99 --group_reporting --iodepth=12 --runtime=3 --ioengine=libaio \
--loops=1 --direct=1 --exitall --name dc --filename=$d | egrep "read|write"
done
done
</TESTSCRIPT>
<4K_RESULTS>
loop driver:
read: IOPS=226k, BW=881MiB/s (924MB/s)(2645MiB/3003msec)
write: IOPS=225k, BW=880MiB/s (923MB/s)(2643MiB/3003msec)
dm-loop target:
read: IOPS=425k, BW=1661MiB/s (1742MB/s)(4990MiB/3004msec)
write: IOPS=425k, BW=1662MiB/s (1743MB/s)(4992MiB/3004msec)
dm-ram target:
read: IOPS=636k, BW=2484MiB/s (2605MB/s)(7464MiB/3005msec)
write: IOPS=636k, BW=2484MiB/s (2605MB/s)(7464MiB/3005msec)
</4K_RESULTS>
<32K_RESULTS>
loop driver:
read: IOPS=55.5k, BW=1733MiB/s (1817MB/s)(5215MiB/3009msec)
write: IOPS=55.2k, BW=1726MiB/s (1810MB/s)(5195MiB/3009msec)
dm-loop target:
read: IOPS=110k, BW=3452MiB/s (3620MB/s)(10.1GiB/3006msec)
write: IOPS=110k, BW=3448MiB/s (3615MB/s)(10.1GiB/3006msec)
dm-ram target:
read: IOPS=355k, BW=10.8GiB/s (11.6GB/s)(32.6GiB/3008msec)
write: IOPS=355k, BW=10.8GiB/s (11.6GB/s)(32.6GiB/3008msec)
</32K_RESULTS>
Signed-off-by: Heinz Mauelshagen <heinzm@xxxxxxxxxx>
Heinz Mauelshagen (2):
dm loop: new target redirecting io to backing file(s)
dm ram: new target redirecting io to RAM
Documentation/device-mapper/loop.txt | 20 ++
Documentation/device-mapper/ram.txt | 15 ++
drivers/md/Kconfig | 14 ++
drivers/md/Makefile | 2 +
drivers/md/dm-loop.c | 352 +++++++++++++++++++++++++++++++++++
drivers/md/dm-ram.c | 101 ++++++++++
6 files changed, 504 insertions(+)
create mode 100644 Documentation/device-mapper/loop.txt
create mode 100644 Documentation/device-mapper/ram.txt
create mode 100644 drivers/md/dm-loop.c
create mode 100644 drivers/md/dm-ram.c
My initial thought for dm-ram was: why? (considering we have brd and
pmem and null_blk). But for 100 lines of code if nothing else it could
serve as yet another example DM target for those interested in learning
more about how to implement a DM target. Would be good to compare its
performance with brd, null_blk and pmem though.
With it we get the dm flexibility to set up ramdisks as opposed to
brd module parameters. It's performance is pretty similar
to brd but it's faster for larger block sizes.
Yes, the value of its simplicity for beginners is an additonal goody.
null_blk doesn't quite fit the list lagging backing store support?
Some numbers in brd, dm-ram, null_blk order:
# fio --bs=32k --rw=randrw --numjobs=99 --group_reporting --iodepth=12
--runtime=3 --ioengine=libaio --loops=1 --direct=1 --exitall --name
pipi --filename=/dev/ram0|egrep "read|write"
read: IOPS=334k, BW=10.2GiB/s (10.0GB/s)(30.7GiB/3009msec)
write: IOPS=334k, BW=10.2GiB/s (10.0GB/s)(30.7GiB/3009msec)
# fio --bs=32k --rw=randrw --numjobs=99 --group_reporting --iodepth=12
--runtime=3 --ioengine=libaio --loops=1 --direct=1 --exitall --name
pipi --filename=/dev/mapper/ram|egrep "read|write"
read: IOPS=354k, BW=10.8GiB/s (11.6GB/s)(32.4GiB/3005msec)
write: IOPS=354k, BW=10.8GiB/s (11.6GB/s)(32.5GiB/3005msec)
# fio --bs=32k --rw=randrw --numjobs=99 --group_reporting --iodepth=12
--runtime=3 --ioengine=libaio --loops=1 --direct=1 --exitall --name
pipi --filename=/dev/nullb0|egrep "read|write"
read: IOPS=337k, BW=10.3GiB/s (11.0GB/s)(30.9GiB/3007msec)
write: IOPS=337k, BW=10.3GiB/s (11.0GB/s)(30.9GiB/3007msec)
As for dm-loop, doubling the performance of the loopback driver is quite
nice (especially with only 1/7 the number of lines of code as
drives/block/loop.c).
Yes, found this challenging in particular too.
Didn't bother to cover direct io or async io (yet).
Much rather wanted to keep it simple.
Cheers,
Heinz
I'll review both of these closer in the coming days but it is getting
_really_ close to the 4.16 merge window (likley opens Sunday) So they
may have to wait until 4.17. We'll see.
Thanks,
Mike
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
