Based on a ML thread by Daniel Berrange we need to test whether this has a similar effect in RHEV when using Extended Page Table (EPT): Tim Bell at CERN published a blog today where he is claiming that they actually see a performance /slowdown/ when EPT is enabled in KVM in their particular workload/benchmarks. "This is a 6% performance improvement with EPT off." http://openstack-in-production.blogspot.co.uk/2015/08/ept-and-ksm-for-high-throughput.html
Please specify what and how to test. As far as I understand this is test-only bug? If so, move to ON_QA, and put the TestOnly keyword.
1. Check that ept is currently on (this is the default): # cat /sys/module/kvm_intel/parameters/ept Y 2. Run a test utilizing cpu and memory and record the results. 3. Make sure you have no running VMs and kvm_intel module not being used. 4. Set a kvm config ept to N: # cat >/etc/modprobe.d/kvm-intel.conf <<EOF options kvm_intel ept=n EOF 5. Reload the kernel module and verify ept state: # modprobe -r kvm_intel # modprobe kvm_intel # cat /sys/module/kvm_intel/parameters/ept N 6. Run the same test you did in step 2 and record the results. 7. Restore ept state (kil VMs, change 'n' to 'y' in step 4, and run step 5). 8. Compare the results from step 2 and 6.
Test performed on RHEVM: 3.6.3-0.1 KVM Version: 2.3.0 - 31.el7_2.7 LIBVIRT Version: libvirt-1.2.17-13.el7_2.3 VDSM Version: vdsm-4.17.19-0.el7ev Test flow: 1. Check ETP settings # cat /sys/module/kvm_intel/parameters/ept Y 2. Populate and start 100 VMs on Hypervisor. Run Load Test using JMeter 100 concurrent threads simulated different actions on 100 VMs which are running on single hypervisor/host during 2 hours Each VM - 1 CPU, 1 G RAM 3. Save monitored statistics 4. Shutdown and remove VMs 4. Update ETL settings # cat >/etc/modprobe.d/kvm-intel.conf <<EOF options kvm_intel ept=n EOF 5. Reload the kernel module and verify ept state: # modprobe -r kvm_intel # modprobe kvm_intel # cat /sys/module/kvm_intel/parameters/ept N 6. Populate and start 100 VMs on Hypervisor and rerun Load Test. 7. Save monitored statistics 8, Analyse results. Test showed us that we have opposite effect when change ETP to N. The better result we reached while EPT is enabled in KVM. Metrics name Module name EPT = Y EPT = N CPU Utilization Server (%) Engine node - System 4.42 4.06 Engine node - User 12.95 6.8 Host – System 7.38 11.78 Host – User 3.38 3.8 CPU Utilization Process (%) Engine node – Ovirt 99.83 46.18 Host node – LIbvirt 19.64 28.04 Host node – VDSM 28.39 30.44 Memory Util Server (Mb) Engine node - Used 27681.43 28608.76 Host - Used 35155.52 34572.57 Memory Util Process (Mb) Engine node – Ovirt 6904.62 6984.05 Host node – LIbvirt 32.36 32.34 Host node – VDSM 542.48 665.1 Network Utilization (Mb) Engine node – Recv 3.71 2.28 Engine node – Sent 3.96 2.41 Host – Recv 57.02 50.21 Host – Sent 1.89 2.21 Disk Utilization (b/sec) Engine node – Reads 0 0 Engine node – Writes 131.33 113.94 Host – Reads 0.02 0.01 Host – Writes 211.16 326.43 See attached detailed document
Created attachment 1132731 [details] Test result
RHEV 3.6 GA was released. This bug didn't have any required doc text, closing.