Performance issue with qcow2/raid

[Jose R. Ziviani]

Hello team,

I'm currently investigating a performance regression detected by iozone 
filesystem benchmark (https://www.iozone.org/).

Basically, if I format a QCOW2 image with XFS filesystem in my guest and run 
iozone I'll get the following result:

$ mkfs.xfs -f /dev/xvdb1 && \
  mount -t xfs /dev/xvdb1 /mnt && \
  /opt/iozone/bin/iozone -a -e -s 16777216 -y 4 -q 8 -i 0 -i 1 -f 
/mnt/iozone.dat

kB        block len  read     reread
16777216  4K         354790   348796
16777216  8K         362356   364818

However, if I revert the commit 46cd1e8a47 (qcow2: Skip copy-on-write when 
allocating a zero cluster) and run the same, I see a huge improvement:

$ mkfs.xfs -f /dev/xvdb1 && \
  mount -t xfs /dev/xvdb1 /mnt && \
  /opt/iozone/bin/iozone -a -e -s 16777216 -y 4 -q 8 -i 0 -i 1 -f 
/mnt/iozone.dat

kB        block len  read     reread
16777216  4K         524067   560057
16777216  8K         538661   537004

Note that if I run iozone without re-formating the disk, I'll get results 
similar to last formatting. In other words, if I my current QEMU executable 
doesn't have commit 46cd1e8a47 and I format the disk, iozone will continue 
showing good results even if I reboot to use QEMU with that commit patched.

My system has a RAID controller[1] and runs QEMU/Xen. I'm not able to reproduce 
such behavior in other systems. 

Do you have any suggestion to help debugging this? What more info could help to 
understand it better?
My next approach is using perf, but I appreciate if you have any hints measure 
qcow efficiently.

[1]
# lspci -vv | grep -i raid
1a:00.0 RAID bus controller: Broadcom / LSI MegaRAID SAS-3 3108 [Invader] (rev 
02)
        Kernel driver in use: megaraid_sas
        Kernel modules: megaraid_sas

Thank you very much!

Jose R. Ziviani

signature.asc
Description: Digital signature