forked from autotest/autotest
-
Notifications
You must be signed in to change notification settings - Fork 0
KVM Autotest Performance Regression Testing
Amos Kong edited this page Dec 22, 2012
·
12 revisions
- Automate all performance testing to save human resource
- Prepare environment automatically, reduce personal error
- Make the result more stable by netperf demo-mode, repeat tests
- Process raw results to a standard format for restoring & comparing easily
- Using statistical method (eg. T-test) to compute average/p-value for comparing results exactly
- iozone (linux & windows ) (iozone has its own result analysis module)
- iometer (windows) (not push upstream)
- ffsb (linux)
- qemu_io (host): (not push upstream)
Autotest already supports prepare environment for performance testing, guest & host need to reboot after setting up.
Autotest supports to numa pining. Assign "numanode=-1" in tests.cfg, then vcpu threads/vhost_net threads/VM memory will be pined to last numa node.
If you want to pin other processes to numa node, you can use numctl and taskset.
memory: numactl -m $n $cmdline cpu: taskset $node_mask $thread_id
You don't need this if you use Autotest framework.
1.First level pinning would be to use numa pinning when starting the guest. e.g numactl -c 1 -m 1 qemu-kvm -smp 2 -m 4G <> (pinning guest memory and cpus to numa-node 1) 2.For a single instance test, it would suggest trying a one to one mapping of vcpu to pyhsical core. e.g get guest vcpu threads id #taskset -p 40 $vcpus1 (pinning vcpu1 thread to pyshical cpu #6 ) #taskset -p 80 $vcpus2 (pinning vcpu2 thread to physical cpu #7 ) 3.To pin vhost on host. get vhost PID and then use taskset to pin it on the same soket. e.g taskset -p 20 $vhost (pinning vcpu2 thread to physical cpu #5 ) 4.In guest,pin the IRQ to one core and the netperf to another. 1) make sure irqbalance is off - `service irqbalance stop` 2) find the interrupts - `cat /proc/interrupts` 3) find the affinity mask for the interrupt(s) - `cat /proc/irq/<irq#>/smp_affinity` 4) change the value to match the proper core.make sure the vlaue is cpu mask. e.g pin the IRQ to first core. echo 01>/proc/irq/$virti0-input/smp_affinity echo 01>/proc/irq/$virti0-output/smp_affinity 5)pin the netserver to another core. e.g taskset -p 02 netserver 5.For host to guest scenario. to get maximum performance. make sure to run netperf on different cores on the same numa node as the guest. e.g numactl -m 1 netperf -T 4 (pinning netperf to physical cpu #4)
- Submit jobs in Autotest server, only execute netperf.guset_exhost for three times.
tests.cfg:
only netperf.guest_exhost
variants:
- repeat1:
- repeat2:
- repeat3:
# vbr0 has a static ip: 192.168.100.16
bridge=vbr0
# virbr0 is created by libvirtd, guest nic2 get ip by dhcp
bridge_nic2 = virbr0
# guest nic1 static ip
ip_nic1 = 192.168.100.21
# external host static ip:
client = 192.168.100.15
Result files:
# cd /usr/local/autotest/results/8-debug_user/192.168.122.1/ # find .|grep RHS kvm.repeat1.r61.virtio_blk.smp2.virtio_net.RHEL.6.1.x86_64.netperf.exhost_guest/results/netperf-result.RHS kvm.repeat2.r61.virtio_blk.smp2.virtio_net.RHEL.6.1.x86_64.netperf.exhost_guest/results/netperf-result.RHS kvm.repeat3.r61.virtio_blk.smp2.virtio_net.RHEL.6.1.x86_64.netperf.exhost_guest/results/netperf-result.RHS
- Submit same job in another env (different packages) with same configuration
Result files:
# cd /usr/local/autotest/results/9-debug_user/192.168.122.1/ # find .|grep RHS kvm.repeat1.r61.virtio_blk.smp2.virtio_net.RHEL.6.1.x86_64.netperf.exhost_guest/results/netperf-result.RHS kvm.repeat2.r61.virtio_blk.smp2.virtio_net.RHEL.6.1.x86_64.netperf.exhost_guest/results/netperf-result.RHS kvm.repeat3.r61.virtio_blk.smp2.virtio_net.RHEL.6.1.x86_64.netperf.exhost_guest/results/netperf-result.RHS
[ntttcp] result_file_pattern = .*.RHS ignore_col = 1 avg_update = [netperf] # testname result_file_pattern = .*.RHS # pattern is used to match result files ignore_col = 2 # some result is the configuration (eg. packet size), we don't need to compute the average of them avg_update = 4,2,3|14,5,12|15,6,13 # update col results after computing averages [iozone] result_file_pattern =
login autotest server by ssh # cd /usr/local/autotest/client/tools 1) compare with log files # python regression.py netperf file /usr/local/autotest/results/8-debug_user/192.168.122.1/ /usr/local/autotest/results/9-debug_user/192.168.122.1/ 2) compare with results in database # python regression.py netperf db 8 9
scipy: http://www.scipy.org/ t-test: http://en.wikipedia.org/wiki/Student's_t-test Two python modules (scipy and numpy) are needed. Script to install numpy/scipy on rhel6 automatically: http://kongove.fedorapeople.org/autotest/scripts/install-numpy-scipy.sh
T-test uses Student's T distribution to compute the probability of difference exists.
Unpaired T-test is used to compare significance of two samples (same configuration), user can check p-value (p-value = 1 - significance) to know if regression bug exists. If the difference of two samples is considered to be not statistically significant(p <= 0.05), it will add a '+' or '-' before p-value. ('+': avg_sample1 < avg_sample2, '-': avg_sample1 > avg_sample2)
Paired t-test is used to compute significance of all averages (all configuration).
- Every Avg line represents the average value based on *$n* repetitions of the same test, and the following SD line represents the Standard Deviation between the *$n* repetitions. - The Standard deviation is displayed as a percentage of the average. - The significance of the differences between the two averages is calculated using unpaired T-test that takes into account the SD of the averages. - The paired t-test is computed for the averages of same category. - only over 95% confidence results will be added "+/-" in "Significance" part. "+" for cpu-usage means regression, "+" for throughput means improvement.
- Raw data that the averages are based on.