which -cpu to use
Hi,
I'm new to KVM and have some questions regarding kvm cpu emulation, hope you can answer them.
I have a kvm node with Dual-Core AMD Opteron(tm) Processor 2218 processors.
1)I was wondering choosing what cpu emulation for VM would give best performance?
I get the following options
# kvm -cpu ? x86 qemu64 x86 phenom x86 core2duo x86 qemu32 x86 coreduo x86 486 x86 pentium x86 pentium2 x86 pentium3 x86 athlon x86 n270 #
I was thinking one of qemu64/phenom/athlon .
Does It matter what 64bit cpu emulation i choose?
What would be chosen if I omit the -cpu option?
2) I want to optimize/recompile a gentoo VM system packages for the chosen cpu - if that matters. Does rebuilding the VM packages to fine tuned for the emulated cpu give any advantages?
3) Do I have to boot a VM kernel with guest viritio drivers, to get a not emulated real cpu access - like I have in Xen?
Thanks a lot.
Alex
Quoted 23 lines · Piavlo
Hi,
I'm new to KVM and have some questions regarding kvm cpu emulation, hope you can answer them.
I have a kvm node with Dual-Core AMD Opteron(tm) Processor 2218 processors.
1)I was wondering choosing what cpu emulation for VM would give best performance?
I get the following options
# kvm -cpu ? x86 qemu64 x86 phenom x86 core2duo x86 qemu32 x86 coreduo x86 486 x86 pentium x86 pentium2 x86 pentium3 x86 athlon x86 n270 #
Quoted 3 lines
I was thinking one of qemu64/phenom/athlon .
Does It matter what 64bit cpu emulation i choose?
What would be chosen if I omit the -cpu option?
If you omit -cpu it chooses the "qemu64" model, which is the best choice for most cases. Don't choose anything else if you don't have any specific reason to do so.
Quoted 2 lines
2) I want to optimize/recompile a gentoo VM system packages for the chosen cpu - if that matters. Does rebuilding the VM packages to fine tuned for the emulated cpu give any advantages?
I don't think features like SSE3 matter that much in a normal environment. You won't get too much more benefit from choosing a different CPU type.
Quoted 1 line
3) Do I have to boot a VM kernel with guest viritio drivers, to get a not emulated real cpu access - like I have in Xen?
virtio drivers have nothing to do with CPU. They are about devices like network and block. If you want to have PV drivers for those, you need to have a kernel that supports virtio (incl. virtio_pci) and specify that you want to use virtio on the qemu command line as options to -net or -drive.
Alex
Quoted 6 lines · Alexander Graf
Quoted 4 lines
2) I want to optimize/recompile a gentoo VM system packages for the chosen cpu - if that matters.
Does rebuilding the VM packages to fine tuned for the emulated cpu give any advantages?I don't think features like SSE3 matter that much in a normal environment.
I just wanted to rebuild the guest with following basic CGLAGS
"-march=opteron -O2 -pipe" (no sse3 and like features) tuned for my hardware node cpu.
In Xen then i use guest linux VM both in fully virtualized mode and paravirtual mode I see in guest's /proc/cpuinfo the same naitive cpu model as in hardware node (while of course some feature flags like svm,sse3 are missing) - but I can build the guests with cpu native gcc -march support anyway. While seeing emulated cpu model in /proc/cpuinfo
model name : QEMU Virtual CPU version 0.9.1
is confusing and i'm not sure if i can use -march=opteron here?
Quoted 5 lines
You won't get too much more benefit from choosing a different CPU type.
Quoted 2 lines
3) Do I have to boot a VM kernel with guest viritio drivers, to get a not emulated real cpu access - like I have in Xen?
virtio drivers have nothing to do with CPU.
Yes I mistakenly used the term "viritio drivers" instead of
"paravirtual guest support". So what I wanted to ask is if I build a guest kernel with paravitual support
will it make the native hardware cpu features available inside the guest like in Xen kernel? Or paravirtual support is for device drivers only and has no impact on CPU handling like in Xen?
I thought that KVM (as Xen) is a bare metal hypervisor with regards to giving access native access to the CPU which have svm or vmx support, and not just CPU emulation.
I'm confused here - can someone shed some light on my ignorance on the matter?
Thanks
Alex
Quoted 6 lines
They are about devices like network and block. If you want to have PV drivers for those, you need to have a kernel that supports virtio (incl.
virtio_pci) and specify that you want to use virtio on the qemu command line as options to -net or -drive.Alex
Quoted 7 lines · Piavlo
Quoted 6 lines · Alexander Graf
Quoted 4 lines
2) I want to optimize/recompile a gentoo VM system packages for the chosen cpu - if that matters.
Does rebuilding the VM packages to fine tuned for the emulated cpu give any advantages?I don't think features like SSE3 matter that much in a normal environment.
Quoted 11 lines
I just wanted to rebuild the guest with following basic CGLAGS "-march=opteron -O2 -pipe" (no sse3 and like features) tuned for my hardware node cpu.
In Xen then i use guest linux VM both in fully virtualized mode and paravirtual mode I see in guest's/proc/cpuinfothe same naitive cpu model as in hardware node (while of course some feature flags like svm,sse3 are missing) - but I can build the guests with cpu native gcc -march support anyway. While seeing emulated cpu model in/proc/cpuinfo
model name : QEMU Virtual CPU version 0.9.1
is confusing and i'm not sure if i can use -march=opteron here?
You can still use all the optimizations you want. -cpu only modifies the CPUID flags exposed to the guest, not the features supported by the CPU. So if your CPU supports say SSE3 and you can't find "pni" in /proc/cpuinfo in the guest, it still works.
Quoted 6 lines
Quoted 6 lines
You won't get too much more benefit from choosing a different CPU type.
Quoted 2 lines
3) Do I have to boot a VM kernel with guest viritio drivers, to get a not emulated real cpu access - like I have in Xen?
virtio drivers have nothing to do with CPU.
Quoted 11 lines
Yes I mistakenly used the term "viritio drivers" instead of "paravirtual guest support". So what I wanted to ask is if I build a guest kernel with paravitual support
will it make the native hardware cpu features available inside the guest like in Xen kernel? Or paravirtual support is for device drivers only and has no impact on CPU handling like in Xen?
I thought that KVM (as Xen) is a bare metal hypervisor with regards to giving access native access to the CPU which have svm or vmx support, and not just CPU emulation.
I'm confused here - can someone shed some light on my ignorance on the matter?
The features are there, you just don't see them in /proc/cpuinfo. Paravirt support is completely unrelated to that.
Alex
Quoted 14 lines · Piavlo
Quoted 1 line · Alexander Graf
virtio drivers have nothing to do with CPU.
Yes I mistakenly used the term "viritio drivers" instead of
"paravirtual guest support". So what I wanted to ask is if I build a guest kernel with paravitual support
will it make the native hardware cpu features available inside the guest like in Xen kernel? Or paravirtual support is for device drivers only and has no impact on CPU handling like in Xen?
I thought that KVM (as Xen) is a bare metal hypervisor with regards to giving access native access to the CPU which have svm or vmx support, and not just CPU emulation.
I'm confused here - can someone shed some light on my ignorance on the matter?
you're mixing several buzzwords:
- paravirtual/fully virtualized guest: this refers as to how the guest runs in terms of the lowest level of hardware, mainly about accessing CPU's ring0, the most privileged mode of execution. paravirtualized guests can't access ring0, so can't do some very low level CPU operations. for this, the guest is modified to use some hypervisor calls instead of CPU operations. since the guest OS has to be modified, usually the VM system doesn't bother fully emulating IO hardware either. Xen works in either mode, KVM only works in 'fully virtualized' mode, using HVM capabilities.
- paravirtualized drivers: usually paravirtualized guests (which KVM doesn't do) also have some special IO channels, which are faster than emulating every detail of an existing piece of hardware. but even if you're fully emulating a CPU, there's nothing preventing you from creating special device drivers that know how to access the Hypervisor communication channels to do any IO needed. the advantage is that these drivers are easy to integrate into an oterwise- unmodified OS. that makes it possible to use PV drivers on windows guests, gaining most (if not all) performance advantages of PV guests without access to the guest OS's source code.
- virtio: is the IO interface presented to the guest for communicating with the emulator. at least KVM uses it, but i think Xen's interface is similar. there are several virtio clients in current Linux kernels, so if you select virtio network, and block device when launching KVM, a Linux guest gets a big speedup. also available if you install the PV drivers on windows guests (for network, block drivers aren't available yet).
- CPU type: this is only how the CPU identifies itself to the guest, and what capabilities it advertises. AFAIK, it doesn't mean any software emulation (á la qemu), or maybe only a few non-performance-sensitive. it's useful mainly to facilitate guest migration between different hosts. if the guest OS sees the same CPU as the host, it might see it changing, and since all modern OSs check the CPU type at bootup to activate different optimised code, changing it while running would make it crash. advertising only the features common to all hosts lets it stay constant no matter how you move the guest around. originally Xen supported migration only between identical hosts, but there's some special features to allow this on some cases. i don't know how complete they're currently.
hope that clears the waters a bit.
Signature · 1 line
Javier
Hi Javier,
thanks for the explanations - they make things much clear to me. Still I have one more misunderstanding - please see below.
Quoted 11 lines · Javier Guerra Giraldez
you're mixing several buzzwords:
- paravirtual/fully virtualized guest: this refers as to how the guest runs in terms of the lowest level of hardware, mainly about accessing CPU's ring0, the most privileged mode of execution. paravirtualized guests can't access ring0, so can't do some very low level CPU operations. for this, the guest is modified to use some hypervisor calls instead of CPU operations. since the guest OS has to be modified, usually the VM system doesn't bother fully emulating IO hardware either. Xen works in either mode, KVM only works in 'fully virtualized' mode, using HVM capabilities.
Quoted 7 lines
- paravirtualized drivers: usually paravirtualized guests (which KVM doesn't do) also have some special IO channels, which are faster than emulating every detail of an existing piece of hardware. but even if you're fully emulating a CPU, there's nothing preventing you from creating special device drivers that know how to access the Hypervisor communication channels to do any IO needed. the advantage is that these drivers are easy to integrate into an oterwise- unmodified OS.
What is still unclear to me is that's the actual difference between PV
drivers implementation in paravirtual
linux guest and PV dirvers in HVM linux guest? AFAIK in xen guest the PV front-end drivers are quite simple, and in KVM guest
to use PV drivers the guest linux needs to be compiled with paravirtual guest and viritio drivers support. So in both cases the OS is modified - but there is the actual difference?
Thanks
Alex
Quoted 26 lines
that makes it possible to use PV drivers on windows guests, gaining most (if not all) performance advantages of PV guests without access to the guest OS's source code.
- virtio: is the IO interface presented to the guest for communicating with the emulator. at least KVM uses it, but i think Xen's interface is similar. there are several virtio clients in current Linux kernels, so if you select virtio network, and block device when launching KVM, a Linux guest gets a big speedup. also available if you install the PV drivers on windows guests (for network, block drivers aren't available yet).
- CPU type: this is only how the CPU identifies itself to the guest, and what capabilities it advertises. AFAIK, it doesn't mean any software emulation (á la qemu), or maybe only a few non-performance-sensitive. it's useful mainly to facilitate guest migration between different hosts. if the guest OS sees the same CPU as the host, it might see it changing, and since all modern OSs check the CPU type at bootup to activate different optimised code, changing it while running would make it crash. advertising only the features common to all hosts lets it stay constant no matter how you move the guest around. originally Xen supported migration only between identical hosts, but there's some special features to allow this on some cases. i don't know how complete they're currently.
hope that clears the waters a bit.
Quoted 7 lines · Piavlo
What is still unclear to me is that's the actual difference between PV drivers implementation in paravirtual
linux guest and PV dirvers in HVM linux guest? AFAIK in xen guest the PV front-end drivers are quite simple, and in KVM guest
to use PV drivers the guest linux needs to be compiled with paravirtual guest and viritio drivers support. So in both cases the OS is modified - but there is the actual difference?
first of all, using paravirtual drivers isn't considered 'core' OS modification. usually they're compiled in the Linux kernel, but could be loaded as modules. and in Windows case, they're always loaded modules. PV guests OTOH, is a very (very) inner-core modification (but a conceptually simple one, the requirements used to be documented in a single HTML for early Xen systems)
now, in KVM (and possibly HVM Xen) the virtio drivers are presented via a special PCI device, while on fully PV guests, there's some function calls that can be used to communicate with the hypervisor. so, there might be some very low-level differences, but above that, they should be the same. as for performance, i really don't think it makes any perceptible difference.
Signature · 1 line
Javier
Ciao Javier Guerra Giraldez,
Quoted 1 line
you're mixing several buzzwords:
thank you for this clear explanation.
Signature · 1 line
Paolo Pedaletti