AW: AW: AW: AW: AW: KVM performance
Hello, I've compiled a new kernel v2.6.27-rc5 with the modified svm.c. But the behaviour of the vlc process in the guest is still the same. I've exported additional cpu features to the guest, e.g. mmxext with kvm-84. But no performance changes. I was not able to export the cpu flags 3dnow and 3dnowext to the guest, no error but they are not visible in /proc/cpuinfo. Not sure why. Regards Stefanie BRAUN, Stefanie wrote: qemu-kvm -cpu ? only shows qemu64, qemu32, 486, pentium, pentium2, pentium3, athlon It can also take additional +feature or -feature parameters. Oh, maybe kvm-84 doesn't have this support? try http://userweb.kernel.org/~avi/kvm-85rc6/. -Ursprüngliche Nachricht- Von: Avi Kivity [mailto:a...@redhat.com] Gesendet: Dienstag, 14. April 2009 10:48 An: BRAUN, Stefanie Cc: kvm@vger.kernel.org Betreff: Re: AW: AW: AW: AW: KVM performance BRAUN, Stefanie wrote: Hello, the host runs on a Dual-Core AMD Opteron Processor. Does there exist a similar AMD parameter? You can add individual host cpu features by using '-cpu qemu64,+feature', where feature is taken from the host /proc/cpuinfo. Do you know which cpu features the program can take advantage of? Also please try replacing the constant 0x0007040600070406ULL in kernel/x86/svm.c with 0x0606060606060606ULL and see what happens (don't forget to reload the modules). -- I have a truly marvellous patch that fixes the bug which this signature is too narrow to contain. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: AW: AW: AW: AW: AW: KVM performance
BRAUN, Stefanie wrote: Hello, I've compiled a new kernel v2.6.27-rc5 with the modified svm.c. But the behaviour of the vlc process in the guest is still the same. I've exported additional cpu features to the guest, e.g. mmxext with kvm-84. But no performance changes. I was not able to export the cpu flags 3dnow and 3dnowext to the guest, no error but they are not visible in /proc/cpuinfo. Not sure why. Can you test on an intel host (relative performance host vs guest)? -- Do not meddle in the internals of kernels, for they are subtle and quick to panic. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
AW: AW: AW: AW: KVM performance
Hello, the host runs on a Dual-Core AMD Opteron Processor. Does there exist a similar AMD parameter? Regard Stefanie -Ursprüngliche Nachricht- Von: Avi Kivity [mailto:a...@redhat.com] Gesendet: Samstag, 11. April 2009 18:19 An: BRAUN, Stefanie Cc: kvm@vger.kernel.org Betreff: Re: AW: AW: AW: KVM performance BRAUN, Stefanie wrote: Hello, now I was able to start the guest vmu with disk virtio, and some of the tests with disk involvement even improved a bit. But the test in which a logo is added to the video stream does not improve. I don't know why the performance is so bad? Subtest: Reading video locally, adding a logo to the video stream and then saving the video locally Host performance: 50% kvm process in host (top) : 99% vlc process in vmu (top) : 99% The output of kvm_stat -1 during the subtest is the following: efer_reload0 0 exits9913473 3994 This indicates that kvm is running in guest mode all of the time and is therefore quite efficient. Perhaps the test uses sse instructions which kvm doesn't expose? Try adding -cpu core2duo to the command line. -- Do not meddle in the internals of kernels, for they are subtle and quick to panic. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: AW: AW: AW: AW: KVM performance
BRAUN, Stefanie wrote: Hello, the host runs on a Dual-Core AMD Opteron Processor. Does there exist a similar AMD parameter? You can add individual host cpu features by using '-cpu qemu64,+feature', where feature is taken from the host /proc/cpuinfo. Do you know which cpu features the program can take advantage of? Also please try replacing the constant 0x0007040600070406ULL in kernel/x86/svm.c with 0x0606060606060606ULL and see what happens (don't forget to reload the modules). -- I have a truly marvellous patch that fixes the bug which this signature is too narrow to contain. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: AW: AW: AW: KVM performance
BRAUN, Stefanie wrote: Hello, now I was able to start the guest vmu with disk virtio, and some of the tests with disk involvement even improved a bit. But the test in which a logo is added to the video stream does not improve. I don't know why the performance is so bad? Subtest: Reading video locally, adding a logo to the video stream and then saving the video locally Host performance: 50% kvm process in host (top) : 99% vlc process in vmu (top) : 99% The output of kvm_stat -1 during the subtest is the following: efer_reload0 0 exits9913473 3994 This indicates that kvm is running in guest mode all of the time and is therefore quite efficient. Perhaps the test uses sse instructions which kvm doesn't expose? Try adding -cpu core2duo to the command line. -- Do not meddle in the internals of kernels, for they are subtle and quick to panic. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
AW: AW: AW: KVM performance
Hello, now I was able to start the guest vmu with disk virtio, and some of the tests with disk involvement even improved a bit. But the test in which a logo is added to the video stream does not improve. I don't know why the performance is so bad? Subtest: Reading video locally, adding a logo to the video stream and then saving the video locally Host performance: 50% kvm process in host (top) : 99% vlc process in vmu (top) : 99% The output of kvm_stat -1 during the subtest is the following: efer_reload0 0 exits9913473 3994 fpu_reload393453 4 halt_exits768222 0 halt_wakeup 497108 0 host_state_reload3266556 4 hypercalls508554 0 insn_emulation 5405007 1999 insn_emulation_fail0 0 invlpg 0 0 io_exits 1879454 0 irq_exits 568541 1995 irq_window 0 0 largepages 0 0 mmio_exits145028 0 mmu_cache_miss 51455 0 mmu_flooded40895 0 mmu_pde_zapped 34101 0 mmu_pte_updated 448719 0 mmu_pte_write 858494 0 mmu_recycled 0 0 mmu_shadow_zapped 50590 0 nmi_window 0 0 pf_fixed 494176 0 pf_guest 378754 0 remote_tlb_flush 0 0 request_irq0 0 signal_exits 1 0 tlb_flush1076949 1 The guest I've started has 512 MB RAM, 1 Core 2.2 GHz of the host which is a dual core machine. Guest settings: RAM: 512 MB CPU: 1 Core with 2,2 GHz Tested i386 OS as well as x86_64 with same performance results. OS: Fedora 10 i386; 2.6.27.21-170.2.56.fc10.i686 CONFIG_HIGHMEM=y CONFIG_HIGHMEM4G=y OS: Fedora 10 x86_64; 2.6.27.21-170.2.56.fc10.x86_64 The config file does not contain the HIGHMEM parameter. Host settings: OS: Fedora 10 x86_64; 2.6.27.5-117.fc10.x86_64 KVM: Version 74 Release 10.fc10 Best regards, Steffi -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
AW: AW: AW: KVM performance
BRAUN, Stefanie wrote: 1. Subtest: VLC reads video from local disk and streams it via udp to another pc Host performance: 11% 11% kvm process in host (top):22% 22% vlc process in vmu (top): 15% 7% While this isn't wonderful, it's not your major bottleneck now. What's the bandwidth generated by the workload? Generated Bandwidth : 6500 kbit per sec 4. Subtest: Reading video locally, adding a logo to the video stream and then saving the video locally Host performance: 50% 50% kvm process in host (top) : 99% 99% vlc process in vmu (top) : 99% 99% Now this is bad. Please provide the output of 'kvm_stat -1' while this is running. Also, describe the guest. Is it Linux? if so, i386 or x86_64? and is CONFIG_HIGHMEM enabled? Linux, Fedora 10, x86_64, (2.6.27.21-170.2.56.fc10.x86_64) The config file does not contain a CONFIG_HIGHMEM parameter. UDP performance is a known issue now, and we are working on it. TCP is much better due to segmentation offload. -- I have a truly marvellous patch that fixes the bug which this signature is too narrow to contain. vmu01_stat Description: vmu01_stat
