I thought that we looked into kvm before and found that it could not be used in our environment, but I can find no evidence of that now. If we can use kvm, it might improve our test run-times.
On the loaner ix box, I followed these instructions: http://developer.android.com/tools/devices/emulator.html https://help.ubuntu.com/community/KVM/Installation Specifically: sudo apt-get install cpu-checker sudo apt-get install qemu-kvm libvirt-bin ubuntu-vm-builder bridge-utils sudo adduser cltbld libvirtd I did not install virt-manager; I did not need to chown anything. To verify installation: [firstname.lastname@example.org ~]$ kvm-ok INFO: /dev/kvm exists KVM acceleration can be used [email@example.com ~]$ virsh -c qemu:///system list Id Name State ---------------------------------- [firstname.lastname@example.org ~]$ ls -l /dev/kvm crw-rw----+ 1 root kvm 10, 232 Oct 18 10:30 /dev/kvm
I added these command line options when starting the emulator: -qemu -m 512 -enable-kvm (this must be at the *end* of the command line). I tried running reftest-1 (of 10 chunks) against nightly. It completed in under 5 minutes!! (Previously only completed 80% of the tests in 60 minutes.) Running full tests suite now...
Over the weekend, I ran the full test suite on the loaner with kvm several times. I found no significant regressions due to kvm: A few extra tests pass and/or fail, so manifests will need an update, but no major changes to results. The run times are fantastic: All reftests and mochitests run much, much faster, while Robocop and xpcshell run at about the same rate. The reftest and mochitest speedup is sufficient to allow us to run R, C, JSR, and M jobs as just one chunk each. However, for R, tests often fail with OOM errors if run as one chunk; they run fine in 2 chunks. Sample run times with kvm: M: 22 minutes M-gl: 2 minutes R1: 23 minutes R2: 22 minutes C: 11 minutes JSR: 5 minutes This is great!
Status: NEW → RESOLVED
Last Resolved: 5 years ago
Resolution: --- → WORKSFORME
You need to log in before you can comment on or make changes to this bug.