[SRU][Xenial][PATCH 0/1] Performance degradation when copying from LVM snapshot backed by NVMe disk

Wed Jun 19 04:34:01 UTC 2019

BugLink: https://bugs.launchpad.net/bugs/1833319

[Impact]
When copying files from a mounted LVM snapshot which resides on NVMe storage 
devices, there is a massive performance degradation in the rate sectors are 
read from the disk.

The kernel is not merging sector requests and is instead issuing many small
sector requests to the NVMe storage controller instead of one larger request.

Experiments have shown a 14x-25x performance degradation in reads, where copies 
used to take seconds, now take minutes, and copies which took thirty minutes 
now take many hours.

The following was found with btrace, running alongside cat (see Testing):

A = IO remapped to different device
Q = IO handled by request queue
G = Get request
U = Unplug request
I = IO inserted onto request queue
D = IO issued to driver
C = IO completion

When reading from the LVM snapshot, we see:

259,0 1 113 0.001117160 1606 A R 837872 + 8 <- (252,0) 835824
259,0 1 114 0.001117276 1606 Q R 837872 + 8 [cat]
259,0 1 115 0.001117451 1606 G R 837872 + 8 [cat]
259,0 1 116 0.001117979 1606 A R 837880 + 8 <- (252,0) 835832
259,0 1 117 0.001118119 1606 Q R 837880 + 8 [cat]
259,0 1 118 0.001118285 1606 G R 837880 + 8 [cat]
259,0 1 122 0.001121613 1606 I RS 837640 + 8 [cat]
259,0 1 123 0.001121687 1606 I RS 837648 + 8 [cat]
259,0 1 124 0.001121758 1606 I RS 837656 + 8 [cat]
...
259,0 1 154 0.001126118 377 D RS 837648 + 8 [kworker/1:1H]
259,0 1 155 0.001126445 377 D RS 837656 + 8 [kworker/1:1H]
259,0 1 156 0.001126871 377 D RS 837664 + 8 [kworker/1:1H]
...
259,0 1 183 0.001848512 0 C RS 837632 + 8 [0]

Now what is happening here, is that a request for 8 sector read is placed onto 
the IO request queue, and is then inserted one at a time to the driver request 
queue and then fetched by the driver.

Comparing this behaviour to reading data from a LVM snapshot on 4.6+ mainline 
or the Ubuntu 4.15 HWE kernel:

M = IO back merged with request on queue

259,0 0 194 0.000532515 1897 A R 7358960 + 8 <- (253,0) 7356912
259,0 0 195 0.000532634 1897 Q R 7358960 + 8 [cat]
259,0 0 196 0.000532810 1897 M R 7358960 + 8 [cat]
259,0 0 197 0.000533864 1897 A R 7358968 + 8 <- (253,0) 7356920
259,0 0 198 0.000533991 1897 Q R 7358968 + 8 [cat]
259,0 0 199 0.000534177 1897 M R 7358968 + 8 [cat]
259,0 0 200 0.000534474 1897 UT N [cat] 1
259,0 0 201 0.000534586 1897 I R 7358464 + 512 [cat]
259,0 0 202 0.000537055 1897 D R 7358464 + 512 [cat]
259,0 0 203 0.002242539 0 C R 7358464 + 512 [0]

This shows us a 8 sector read is added to the request queue, and is then
subsequently [M]erged backward with other requests on the queue until the sum 
of all of those merged requests becomes 512 sectors. From there, the 512 sector 
read is placed onto the IO queue, where it is fetched by the device driver, and 
completes.

[Fix]

The problem is that the NVMe driver on 4.4 xenial kernel is not merging 8 
sector requests.

Merging is controlled per device by this sysfs entry:
/sys/block/nvme1n1/queue/nomerges

On 4.4 xenial, reading from this yields 2, or (QUEUE_FLAG_NOMERGES).
On 4.5+ and 4.15 HWE kernel, reading from this yields 0, or allowing merge.

Setting this to 0 on the 4.4 kernel with:

# echo "0" > /sys/block/nvme1n1/queue/nomerges

and testing again, we find performance is restored and the problem is fixed.

Performing a btrace, we see 8 sector reads get backmerged into a 512 sector 
read which is done in one go.

The problem was fixed in 4.5 upstream with the below commit:

commit ef2d4615c59efb312e531a5e949970f37ca1c841
Author: Keith Busch <keith.busch at intel.com>
Date: Thu Feb 11 13:05:40 2016 -0700
Subject: NVMe: Allow request merges

This commit removes the QUEUE_FLAG_NOMERGES flag from being set during driver 
init, allowing requests to be backmerged. This also has a direct effect of 
defaulting /sys/block/nvme1n1/queue/nomerges to 0.

Please cherry-pick ef2d4615c59efb312e531a5e949970f37ca1c841 to all xenial 4.4
kernels.

[Testcase]

You can replicate the problem with a system with a NVMe disk. I recommend using 
c5.large AWS EC2 instances with a secondary gpt2 EBS disk of 200gb or larger.

Steps (with NVMe disk being /dev/nvme1n1):
  1. sudo pvcreate /dev/nvme1n1
  2. sudo vgcreate secvol /dev/nvme1n1
  3. sudo lvcreate --name seclv -l 80%FREE secvol
  4. sudo mkfs.ext4 /dev/secvol/seclv
  5. sudo mount /dev/mapper/secvol-seclv /mnt
  6. for i in `seq 1 20`; do sudo dd if=/dev/zero of=/mnt/dummy$i bs=512M count=1; done
  7. sudo lvcreate --snapshot /dev/secvol/seclv --name tmp_backup1 --extents '90%FREE'
  8. NEWMOUNT=$(mktemp -t -d mount.backup_XXX)
  9. sudo mount -v -o ro /dev/secvol/tmp_backup1 $NEWMOUNT

To replicate, simply read one of those 512mb files:
  10. time cat $NEWMOUNT/dummy1 1> /dev/null

On a stock xenial kernel, expect to see the following:

4.4.0-151-generic #178-Ubuntu

$ time cat /tmp/mount.backup_TYD/dummy1 1> /dev/null

real 0m42.693s
user 0m0.008s
sys 0m0.388s
$ cat /sys/block/nvme1n1/queue/nomerges
2

On a patched xenial kernel, performance is restored:

4.4.0-151-generic #178+hf228435v20190618b1-Ubuntu

$ time cat /tmp/mount.backup_aId/dummy1 1> /dev/null

real 0m1.773s
user 0m0.008s
sys 0m0.184s
$ cat /sys/block/nvme1n1/queue/nomerges
0

[Regression Potential]

Cherry-picking "NVMe: Allow request merges" changes the default request policy 
for NVMe drives, which may give some cause for concern in both terms of 
stability and performance for other workloads.

Regarding stability, this flag was originally set when the NVMe driver was
bio based, before the driver had been converted to blk-mq and separated out 
from /block. You can read a mailing list thread about it here:

https://lists.infradead.org/pipermail/linux-nvme/2016-February/003946.html

Along with the commit "MD: make bio mergeable" there is no reason to not allow 
requests to be mergeable for the new NVMe driver.

Regarding performance for other workloads, I reference the commit which
QUEUE_FLAG_NOMERGES or nomerges == 2 was introduced:
commit: 488991e28e55b4fbca8067edf0259f69d1a6f92c
subject: block: Added in stricter no merge semantics for block I/O

nomerges Throughput %System Improvement (tput / %sys)
-------- ------------ ----------- -------------------------
0 12.45 MB/sec 0.669365609
1 12.50 MB/sec 0.641519199 0.40% / 2.71%
2 12.52 MB/sec 0.639849750 0.56% / 2.96%

It shows a 0.56% performance increase for no merging / 2, over allowing
merging / 0 for random IO workloads.

Comparing this with the 14x-25x performance degradation for reads where requests
are not able to be merged, it is clear that changing the default to 0 will not
impact any other workloads by any significant margin.

The commit is also present in Linux 4.5 mainline, can be cleanly cherry-picked 
and is still present in the kernel to this day, and after review of the NVMe 
driver, I believe there will be no regressions.

If you are interested in testing, I have prepared two ppas with
ef2d4615c59efb312e531a5e949970f37ca1c841 patched:

linux-image-generic: https://launchpad.net/~mruffell/+archive/ubuntu/sf228435-test-generic
linux-image-aws: https://launchpad.net/~mruffell/+archive/ubuntu/sf228435-test

Keith Busch (1):
  NVMe: Allow request merges

 drivers/nvme/host/core.c | 1 -
 1 file changed, 1 deletion(-)

-- 
2.20.1