[PATCH] v2v: Optimize convert for images with small holes
"qemu-img convert" detects zeroes in allocated areas and punch holes in the destination image. This may save space on the destination image, but slows down conversion when using outputs such as rhv-upload, which have very large overhead per requests. Using the -S flag, we can treat small areas filled with zeroes as data, limiting the number of requests, and speeding the operation. Here is an example converting Fedora 30 image: $ virt-builder fedora-30 -o src.img ... $ qemu-img map -f raw --output json src.img | wc -l 213 $ qemu-img convert -f raw -O raw -t none -T none src.img dst.img $ qemu-img map -f raw --output json dst.img | wc -l 1443 $ ls -lhs *.img 1.2G -rw-r--r--. 1 nsoffer nsoffer 6.0G Nov 1 21:48 dst.img 1.2G -rw-r--r--. 1 nsoffer nsoffer 6.0G Nov 1 21:46 src.img Qemu did 1443 writes instead of 213 (5.8X). Lets repeat this conversion with the -S option: $ qemu-img convert -f raw -O raw -t none -T none -S 64k src.img dst.img $ qemu-img map -f raw --output json dst.img | wc -l 213 $ ls -lhs *.img 1.2G -rw-r--r--. 1 nsoffer nsoffer 6.0G Nov 1 21:48 dst.img 1.2G -rw-r--r--. 1 nsoffer nsoffer 6.0G Nov 1 21:46 src.img Picking a good value for -S is not easy. Testing show that 64k is best value for this test image for limiting the number of requests: $ for size in 4k 8k 16k 32k 64k; do \ printf "%5s: " $size; \ qemu-img convert -f raw -O raw -t none -T none -S $size src.img dst.img; \ qemu-img map -f raw --output json dst.img | wc -l; \ done 4k: 1443 8k: 731 16k: 521 32k: 387 64k: 213 We need more testing with oVirt to measure the performance improvement and pick a good value. This should probably be an option, but lets start with a minimal change. --- Untested. I cannot build virt-v2v on Fedora 29 since libguestfs 1.41.5 is required but not available. I can also not build libguestfs from git becuase setting up common submodule in ./autogen.sh fails. v2v/v2v.ml | 1 + 1 file changed, 1 insertion(+) diff --git a/v2v/v2v.ml b/v2v/v2v.ml index 4655b883..03590c9e 100644 --- a/v2v/v2v.ml +++ b/v2v/v2v.ml @@ -741,6 +741,7 @@ and copy_targets cmdline targets input output = (if not (quiet ()) then [ "-p" ] else []) @ [ "-n"; "-f"; "qcow2"; "-O"; t.target_format ] @ (if cmdline.compressed then [ "-c" ] else []) @ + [ "-S"; "64k" ] @ [ overlay_file; filename ] in let start_time = gettimeofday () in if run_command cmd <> 0 then -- 2.20.1
On Fri, Nov 1, 2019, 22:56 Nir Soffer <nirsof@gmail.com> wrote:
"qemu-img convert" detects zeroes in allocated areas and punch holes in the destination image. This may save space on the destination image, but slows down conversion when using outputs such as rhv-upload, which have very large overhead per requests.
Using the -S flag, we can treat small areas filled with zeroes as data, limiting the number of requests, and speeding the operation.
Here is an example converting Fedora 30 image:
$ virt-builder fedora-30 -o src.img ...
$ qemu-img map -f raw --output json src.img | wc -l 213
$ qemu-img convert -f raw -O raw -t none -T none src.img dst.img
$ qemu-img map -f raw --output json dst.img | wc -l 1443
$ ls -lhs *.img 1.2G -rw-r--r--. 1 nsoffer nsoffer 6.0G Nov 1 21:48 dst.img 1.2G -rw-r--r--. 1 nsoffer nsoffer 6.0G Nov 1 21:46 src.img
Qemu did 1443 writes instead of 213 (5.8X). Lets repeat this conversion with the -S option:
$ qemu-img convert -f raw -O raw -t none -T none -S 64k src.img dst.img
$ qemu-img map -f raw --output json dst.img | wc -l 213
$ ls -lhs *.img 1.2G -rw-r--r--. 1 nsoffer nsoffer 6.0G Nov 1 21:48 dst.img 1.2G -rw-r--r--. 1 nsoffer nsoffer 6.0G Nov 1 21:46 src.img
Picking a good value for -S is not easy. Testing show that 64k is best value for this test image for limiting the number of requests:
$ for size in 4k 8k 16k 32k 64k; do \ printf "%5s: " $size; \ qemu-img convert -f raw -O raw -t none -T none -S $size src.img dst.img; \ qemu-img map -f raw --output json dst.img | wc -l; \ done 4k: 1443 8k: 731 16k: 521 32k: 387 64k: 213
We need more testing with oVirt to measure the performance improvement and pick a good value. This should probably be an option, but lets start with a minimal change. ---
Untested. I cannot build virt-v2v on Fedora 29 since libguestfs 1.41.5 is required but not available. I can also not build libguestfs from git becuase setting up common submodule in ./autogen.sh fails.
v2v/v2v.ml | 1 + 1 file changed, 1 insertion(+)
diff --git a/v2v/v2v.ml b/v2v/v2v.ml index 4655b883..03590c9e 100644 --- a/v2v/v2v.ml +++ b/v2v/v2v.ml @@ -741,6 +741,7 @@ and copy_targets cmdline targets input output = (if not (quiet ()) then [ "-p" ] else []) @ [ "-n"; "-f"; "qcow2"; "-O"; t.target_format ] @ (if cmdline.compressed then [ "-c" ] else []) @ + [ "-S"; "64k" ] @ [ overlay_file; filename ] in let start_time = gettimeofday () in if run_command cmd <> 0 then
I finally built this by applying the patch to libguestfs 1.40.1. I did initial testing with local output like this: $ ./run virt-v2v -v -x -i disk /home/nsoffer/tmp/sparse-size/src.img -o local -of raw -os /home/nsoffer/tmp/sparse-size Testing show that this change improves throughput, but converting with local file system is so fast that this is not interesting. I hope to more interesting results with rhv-upload. Hopefully I can test this next week. --- With 1.40.1: qemu-img 'convert' '-p' '-n' '-f' 'qcow2' '-O' 'raw' '/home/nsoffer/src/libguestfs/tmp/v2vovld6b8b9.qcow2' '/home/nsoffer/tmp/sparse-size/src-sda' (100.00/100%) du --block-size=1 '/home/nsoffer/tmp/sparse-size/src-sda' | awk '{print $1}' virtual copying rate: 59718.8 M bits/sec real copying rate: 11479.2 M bits/sec sda: estimate 2456662831 (2.3G) versus actual 1238372352 (1.2G): 98.4% qemu-img 'convert' '-p' '-n' '-f' 'qcow2' '-O' 'raw' '/home/nsoffer/src/libguestfs/tmp/v2vovl70b9b6.qcow2' '/home/nsoffer/tmp/sparse-size/src-sda' (100.00/100%) du --block-size=1 '/home/nsoffer/tmp/sparse-size/src-sda' | awk '{print $1}' virtual copying rate: 65311.3 M bits/sec real copying rate: 12567.4 M bits/sec sda: estimate 2456662831 (2.3G) versus actual 1239678976 (1.2G): 98.2% qemu-img 'convert' '-p' '-n' '-f' 'qcow2' '-O' 'raw' '/home/nsoffer/src/libguestfs/tmp/v2vovl1c8ada.qcow2' '/home/nsoffer/tmp/sparse-size/src-sda' (100.00/100%) du --block-size=1 '/home/nsoffer/tmp/sparse-size/src-sda' | awk '{print $1}' virtual copying rate: 68175.6 M bits/sec real copying rate: 13118.6 M bits/sec sda: estimate 2456662831 (2.3G) versus actual 1239678976 (1.2G): 98.2% With 1.40.1 + patch: qemu-img 'convert' '-p' '-n' '-f' 'qcow2' '-O' 'raw' '-S' '64k' '/home/nsoffer/src/libguestfs/tmp/v2vovl981419.qcow2' '/home/nsoffer/tmp/sparse-size/src-sda' (100.00/100%) du --block-size=1 '/home/nsoffer/tmp/sparse-size/src-sda' | awk '{print $1}' virtual copying rate: 70717.5 M bits/sec real copying rate: 13763.8 M bits/sec sda: estimate 2456662831 (2.3G) versus actual 1253900288 (1.2G): 95.9% qemu-img 'convert' '-p' '-n' '-f' 'qcow2' '-O' 'raw' '-S' '64k' '/home/nsoffer/src/libguestfs/tmp/v2vovl10e7e3.qcow2' '/home/nsoffer/tmp/sparse-size/src-sda' (100.00/100%) du --block-size=1 '/home/nsoffer/tmp/sparse-size/src-sda' | awk '{print $1}' virtual copying rate: 73361.8 M bits/sec real copying rate: 14278.5 M bits/sec sda: estimate 2456662831 (2.3G) versus actual 1253900288 (1.2G): 95.9% qemu-img 'convert' '-p' '-n' '-f' 'qcow2' '-O' 'raw' '-S' '64k' '/home/nsoffer/src/libguestfs/tmp/v2vovldf5aaf.qcow2' '/home/nsoffer/tmp/sparse-size/src-sda' (100.00/100%) du --block-size=1 '/home/nsoffer/tmp/sparse-size/src-sda' | awk '{print $1}' virtual copying rate: 73106.5 M bits/sec real copying rate: 14228.8 M bits/sec sda: estimate 2456662831 (2.3G) versus actual 1253900288 (1.2G): 95.9%
On Fri, Nov 01, 2019 at 10:56:18PM +0100, Nir Soffer wrote:
Untested. I cannot build virt-v2v on Fedora 29 since libguestfs 1.41.5 is required but not available. I can also not build libguestfs from git becuase setting up common submodule in ./autogen.sh fails.
You can try doing: git submodule init git submodule update Note if there's a common/ subdirectory left over with temporary files in it, you have to delete it first. Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-df lists disk usage of guests without needing to install any software inside the virtual machine. Supports Linux and Windows. http://people.redhat.com/~rjones/virt-df/
On Fri, Nov 01, 2019 at 10:56:18PM +0100, Nir Soffer wrote:
diff --git a/v2v/v2v.ml b/v2v/v2v.ml index 4655b883..03590c9e 100644 --- a/v2v/v2v.ml +++ b/v2v/v2v.ml @@ -741,6 +741,7 @@ and copy_targets cmdline targets input output = (if not (quiet ()) then [ "-p" ] else []) @ [ "-n"; "-f"; "qcow2"; "-O"; t.target_format ] @ (if cmdline.compressed then [ "-c" ] else []) @ + [ "-S"; "64k" ] @
I kind of wonder if we should make this something that the output mode can set. However I don't think it's a big deal since (a) 64K is the same as the standard qcow2 cluster size so it won't affect qcow2 outputs and (b) it's not a particularly large size even for raw outputs. Therefore: ACK Rich. -- Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones Read my programming and virtualization blog: http://rwmj.wordpress.com virt-top is 'top' for virtual machines. Tiny program with many powerful monitoring features, net stats, disk stats, logging, etc. http://people.redhat.com/~rjones/virt-top
participants (3)
-
Nir Soffer -
Nir Soffer -
Richard W.M. Jones