Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
On 14/11/17 12:43, Quan Xu wrote: > > > On 2017/11/14 18:27, Juergen Gross wrote: >> On 14/11/17 10:38, Quan Xu wrote: >>> >>> On 2017/11/14 15:30, Juergen Gross wrote: On 14/11/17 08:02, Quan Xu wrote: > On 2017/11/13 18:53, Juergen Gross wrote: >> On 13/11/17 11:06, Quan Xu wrote: >>> From: Quan Xu >>> >>> So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is >>> called >>> in idle path which will poll for a while before we enter the real >>> idle >>> state. >>> >>> In virtualization, idle path includes several heavy operations >>> includes timer access(LAPIC timer or TSC deadline timer) which will >>> hurt performance especially for latency intensive workload like >>> message >>> passing task. The cost is mainly from the vmexit which is a hardware >>> context switch between virtual machine and hypervisor. Our >>> solution is >>> to poll for a while and do not enter real idle path if we can get >>> the >>> schedule event during polling. >>> >>> Poll may cause the CPU waste so we adopt a smart polling >>> mechanism to >>> reduce the useless poll. >>> >>> Signed-off-by: Yang Zhang >>> Signed-off-by: Quan Xu >>> Cc: Juergen Gross >>> Cc: Alok Kataria >>> Cc: Rusty Russell >>> Cc: Thomas Gleixner >>> Cc: Ingo Molnar >>> Cc: "H. Peter Anvin" >>> Cc: x...@kernel.org >>> Cc: virtualizat...@lists.linux-foundation.org >>> Cc: linux-kernel@vger.kernel.org >>> Cc: xen-de...@lists.xenproject.org >> Hmm, is the idle entry path really so critical to performance that a >> new >> pvops function is necessary? > Juergen, Here is the data we get when running benchmark netperf: > 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): > 29031.6 bit/s -- 76.1 %CPU > > 2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0): > 35787.7 bit/s -- 129.4 %CPU > > 3. w/ kvm dynamic poll: > 35735.6 bit/s -- 200.0 %CPU > > 4. w/patch and w/ kvm dynamic poll: > 42225.3 bit/s -- 198.7 %CPU > > 5. idle=poll > 37081.7 bit/s -- 998.1 %CPU > > > > w/ this patch, we will improve performance by 23%.. even we could > improve > performance by 45.4%, if we use w/patch and w/ kvm dynamic poll. > also the > cost of CPU is much lower than 'idle=poll' case.. I don't question the general idea. I just think pvops isn't the best way to implement it. >> Wouldn't a function pointer, maybe guarded >> by a static key, be enough? A further advantage would be that this >> would >> work on other architectures, too. > I assume this feature will be ported to other archs.. a new pvops > makes >>> sorry, a typo.. /other archs/other hypervisors/ >>> it refers hypervisor like Xen, HyperV and VMware).. >>> > code > clean and easy to maintain. also I tried to add it into existed pvops, > but it > doesn't match. You are aware that pvops is x86 only? >>> yes, I'm aware.. >>> I really don't see the big difference in maintainability compared to the static key / function pointer variant: void (*guest_idle_poll_func)(void); struct static_key guest_idle_poll_key __read_mostly; static inline void guest_idle_poll(void) { if (static_key_false(&guest_idle_poll_key)) guest_idle_poll_func(); } >>> >>> >>> thank you for your sample code :) >>> I agree there is no big difference.. I think we are discussion for two >>> things: >>> 1) x86 VM on different hypervisors >>> 2) different archs VM on kvm hypervisor >>> >>> What I want to do is x86 VM on different hypervisors, such as kvm / xen >>> / hyperv .. >> Why limit the solution to x86 if the more general solution isn't >> harder? >> >> As you didn't give any reason why the pvops approach is better other >> than you don't care for non-x86 platforms you won't get an "Ack" from >> me for this patch. > > > It just looks a little odder to me. I understand you care about no-x86 > arch. > > Are you aware 'pv_time_ops' for arm64/arm/x86 archs, defined in > - arch/arm64/include/asm/paravirt.h > - arch/x86/include/asm/paravirt_types.h > - arch/arm/include/asm/paravirt.h Yes, I know. This is just a hack to make it compile. Other than the same names this has nothing to do with pvops, but is just a function vector. > I am unfamilar with arm code. IIUC, if you'd implement pv_idle_ops > for arm/arm64 arch, you'd define a same structure in > - arch/arm64/include/asm/paravirt.h or > - arch/arm/include/asm/paravirt.h > > .. instead of static key / fuction. > > then implement a real function in > - arch/arm/kernel/paravirt.c. So just to use pvops you want to implement it in each arch instead of using a mechanism available everywhere? >
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
On 2017/11/14 18:27, Juergen Gross wrote: On 14/11/17 10:38, Quan Xu wrote: On 2017/11/14 15:30, Juergen Gross wrote: On 14/11/17 08:02, Quan Xu wrote: On 2017/11/13 18:53, Juergen Gross wrote: On 13/11/17 11:06, Quan Xu wrote: From: Quan Xu So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called in idle path which will poll for a while before we enter the real idle state. In virtualization, idle path includes several heavy operations includes timer access(LAPIC timer or TSC deadline timer) which will hurt performance especially for latency intensive workload like message passing task. The cost is mainly from the vmexit which is a hardware context switch between virtual machine and hypervisor. Our solution is to poll for a while and do not enter real idle path if we can get the schedule event during polling. Poll may cause the CPU waste so we adopt a smart polling mechanism to reduce the useless poll. Signed-off-by: Yang Zhang Signed-off-by: Quan Xu Cc: Juergen Gross Cc: Alok Kataria Cc: Rusty Russell Cc: Thomas Gleixner Cc: Ingo Molnar Cc: "H. Peter Anvin" Cc: x...@kernel.org Cc: virtualizat...@lists.linux-foundation.org Cc: linux-kernel@vger.kernel.org Cc: xen-de...@lists.xenproject.org Hmm, is the idle entry path really so critical to performance that a new pvops function is necessary? Juergen, Here is the data we get when running benchmark netperf: 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): 29031.6 bit/s -- 76.1 %CPU 2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0): 35787.7 bit/s -- 129.4 %CPU 3. w/ kvm dynamic poll: 35735.6 bit/s -- 200.0 %CPU 4. w/patch and w/ kvm dynamic poll: 42225.3 bit/s -- 198.7 %CPU 5. idle=poll 37081.7 bit/s -- 998.1 %CPU w/ this patch, we will improve performance by 23%.. even we could improve performance by 45.4%, if we use w/patch and w/ kvm dynamic poll. also the cost of CPU is much lower than 'idle=poll' case.. I don't question the general idea. I just think pvops isn't the best way to implement it. Wouldn't a function pointer, maybe guarded by a static key, be enough? A further advantage would be that this would work on other architectures, too. I assume this feature will be ported to other archs.. a new pvops makes sorry, a typo.. /other archs/other hypervisors/ it refers hypervisor like Xen, HyperV and VMware).. code clean and easy to maintain. also I tried to add it into existed pvops, but it doesn't match. You are aware that pvops is x86 only? yes, I'm aware.. I really don't see the big difference in maintainability compared to the static key / function pointer variant: void (*guest_idle_poll_func)(void); struct static_key guest_idle_poll_key __read_mostly; static inline void guest_idle_poll(void) { if (static_key_false(&guest_idle_poll_key)) guest_idle_poll_func(); } thank you for your sample code :) I agree there is no big difference.. I think we are discussion for two things: 1) x86 VM on different hypervisors 2) different archs VM on kvm hypervisor What I want to do is x86 VM on different hypervisors, such as kvm / xen / hyperv .. Why limit the solution to x86 if the more general solution isn't harder? As you didn't give any reason why the pvops approach is better other than you don't care for non-x86 platforms you won't get an "Ack" from me for this patch. It just looks a little odder to me. I understand you care about no-x86 arch. Are you aware 'pv_time_ops' for arm64/arm/x86 archs, defined in - arch/arm64/include/asm/paravirt.h - arch/x86/include/asm/paravirt_types.h - arch/arm/include/asm/paravirt.h I am unfamilar with arm code. IIUC, if you'd implement pv_idle_ops for arm/arm64 arch, you'd define a same structure in - arch/arm64/include/asm/paravirt.h or - arch/arm/include/asm/paravirt.h .. instead of static key / fuction. then implement a real function in - arch/arm/kernel/paravirt.c. Also I wonder HOW/WHERE to define a static key/function, then to benifit x86/no-x86 archs? Quan Alibaba Cloud And KVM would just need to set guest_idle_poll_func and enable the static key. Works on non-x86 architectures, too. .. referred to 'pv_mmu_ops', HyperV and Xen can implement their own functions for 'pv_mmu_ops'. I think it is the same to pv_idle_ops. with above explaination, do you still think I need to define the static key/function pointer variant? btw, any interest to port it to Xen HVM guest? :) Maybe. But this should work for Xen on ARM, too. Juergen
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
On 14/11/17 10:38, Quan Xu wrote: > > > On 2017/11/14 15:30, Juergen Gross wrote: >> On 14/11/17 08:02, Quan Xu wrote: >>> >>> On 2017/11/13 18:53, Juergen Gross wrote: On 13/11/17 11:06, Quan Xu wrote: > From: Quan Xu > > So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called > in idle path which will poll for a while before we enter the real idle > state. > > In virtualization, idle path includes several heavy operations > includes timer access(LAPIC timer or TSC deadline timer) which will > hurt performance especially for latency intensive workload like > message > passing task. The cost is mainly from the vmexit which is a hardware > context switch between virtual machine and hypervisor. Our solution is > to poll for a while and do not enter real idle path if we can get the > schedule event during polling. > > Poll may cause the CPU waste so we adopt a smart polling mechanism to > reduce the useless poll. > > Signed-off-by: Yang Zhang > Signed-off-by: Quan Xu > Cc: Juergen Gross > Cc: Alok Kataria > Cc: Rusty Russell > Cc: Thomas Gleixner > Cc: Ingo Molnar > Cc: "H. Peter Anvin" > Cc: x...@kernel.org > Cc: virtualizat...@lists.linux-foundation.org > Cc: linux-kernel@vger.kernel.org > Cc: xen-de...@lists.xenproject.org Hmm, is the idle entry path really so critical to performance that a new pvops function is necessary? >>> Juergen, Here is the data we get when running benchmark netperf: >>> 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): >>> 29031.6 bit/s -- 76.1 %CPU >>> >>> 2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0): >>> 35787.7 bit/s -- 129.4 %CPU >>> >>> 3. w/ kvm dynamic poll: >>> 35735.6 bit/s -- 200.0 %CPU >>> >>> 4. w/patch and w/ kvm dynamic poll: >>> 42225.3 bit/s -- 198.7 %CPU >>> >>> 5. idle=poll >>> 37081.7 bit/s -- 998.1 %CPU >>> >>> >>> >>> w/ this patch, we will improve performance by 23%.. even we could >>> improve >>> performance by 45.4%, if we use w/patch and w/ kvm dynamic poll. >>> also the >>> cost of CPU is much lower than 'idle=poll' case.. >> I don't question the general idea. I just think pvops isn't the best way >> to implement it. >> Wouldn't a function pointer, maybe guarded by a static key, be enough? A further advantage would be that this would work on other architectures, too. >>> I assume this feature will be ported to other archs.. a new pvops makes > > sorry, a typo.. /other archs/other hypervisors/ > it refers hypervisor like Xen, HyperV and VMware).. > >>> code >>> clean and easy to maintain. also I tried to add it into existed pvops, >>> but it >>> doesn't match. >> You are aware that pvops is x86 only? > > yes, I'm aware.. > >> I really don't see the big difference in maintainability compared to the >> static key / function pointer variant: >> >> void (*guest_idle_poll_func)(void); >> struct static_key guest_idle_poll_key __read_mostly; >> >> static inline void guest_idle_poll(void) >> { >> if (static_key_false(&guest_idle_poll_key)) >> guest_idle_poll_func(); >> } > > > > thank you for your sample code :) > I agree there is no big difference.. I think we are discussion for two > things: > 1) x86 VM on different hypervisors > 2) different archs VM on kvm hypervisor > > What I want to do is x86 VM on different hypervisors, such as kvm / xen > / hyperv .. Why limit the solution to x86 if the more general solution isn't harder? As you didn't give any reason why the pvops approach is better other than you don't care for non-x86 platforms you won't get an "Ack" from me for this patch. > >> And KVM would just need to set guest_idle_poll_func and enable the >> static key. Works on non-x86 architectures, too. >> > > .. referred to 'pv_mmu_ops', HyperV and Xen can implement their own > functions for 'pv_mmu_ops'. > I think it is the same to pv_idle_ops. > > with above explaination, do you still think I need to define the static > key/function pointer variant? > > btw, any interest to port it to Xen HVM guest? :) Maybe. But this should work for Xen on ARM, too. Juergen
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
On 2017/11/14 16:22, Wanpeng Li wrote: 2017-11-14 16:15 GMT+08:00 Quan Xu : On 2017/11/14 15:12, Wanpeng Li wrote: 2017-11-14 15:02 GMT+08:00 Quan Xu : On 2017/11/13 18:53, Juergen Gross wrote: On 13/11/17 11:06, Quan Xu wrote: From: Quan Xu So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called in idle path which will poll for a while before we enter the real idle state. In virtualization, idle path includes several heavy operations includes timer access(LAPIC timer or TSC deadline timer) which will hurt performance especially for latency intensive workload like message passing task. The cost is mainly from the vmexit which is a hardware context switch between virtual machine and hypervisor. Our solution is to poll for a while and do not enter real idle path if we can get the schedule event during polling. Poll may cause the CPU waste so we adopt a smart polling mechanism to reduce the useless poll. Signed-off-by: Yang Zhang Signed-off-by: Quan Xu Cc: Juergen Gross Cc: Alok Kataria Cc: Rusty Russell Cc: Thomas Gleixner Cc: Ingo Molnar Cc: "H. Peter Anvin" Cc: x...@kernel.org Cc: virtualizat...@lists.linux-foundation.org Cc: linux-kernel@vger.kernel.org Cc: xen-de...@lists.xenproject.org Hmm, is the idle entry path really so critical to performance that a new pvops function is necessary? Juergen, Here is the data we get when running benchmark netperf: 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): 29031.6 bit/s -- 76.1 %CPU 2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0): 35787.7 bit/s -- 129.4 %CPU 3. w/ kvm dynamic poll: 35735.6 bit/s -- 200.0 %CPU Actually we can reduce the CPU utilization by sleeping a period of time as what has already been done in the poll logic of IO subsystem, then we can improve the algorithm in kvm instead of introduing another duplicate one in the kvm guest. We really appreciate upstream's kvm dynamic poll mechanism, which is really helpful for a lot of scenario.. However, as description said, in virtualization, idle path includes several heavy operations includes timer access (LAPIC timer or TSC deadline timer) which will hurt performance especially for latency intensive workload like message passing task. The cost is mainly from the vmexit which is a hardware context switch between virtual machine and hypervisor. for upstream's kvm dynamic poll mechanism, even you could provide a better algorism, how could you bypass timer access (LAPIC timer or TSC deadline timer), or a hardware context switch between virtual machine and hypervisor. I know these is a tradeoff. Furthermore, here is the data we get when running benchmark contextswitch to measure the latency(lower is better): 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): 3402.9 ns/ctxsw -- 199.8 %CPU 2. w/ patch and disable kvm dynamic poll: 1163.5 ns/ctxsw -- 205.5 %CPU 3. w/ kvm dynamic poll: 2280.6 ns/ctxsw -- 199.5 %CPU so, these tow solution are quite similar, but not duplicate.. that's also why to add a generic idle poll before enter real idle path. When a reschedule event is pending, we can bypass the real idle path. There is a similar logic in the idle governor/driver, so how this patchset influence the decision in the idle governor/driver when running on bare-metal(power managment is not exposed to the guest so we will not enter into idle driver in the guest)? This is expected to take effect only when running as a virtual machine with proper CONFIG_* enabled. This can not work on bare mental even with proper CONFIG_* enabled. Quan Alibaba Cloud
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
On 2017/11/14 15:30, Juergen Gross wrote: On 14/11/17 08:02, Quan Xu wrote: On 2017/11/13 18:53, Juergen Gross wrote: On 13/11/17 11:06, Quan Xu wrote: From: Quan Xu So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called in idle path which will poll for a while before we enter the real idle state. In virtualization, idle path includes several heavy operations includes timer access(LAPIC timer or TSC deadline timer) which will hurt performance especially for latency intensive workload like message passing task. The cost is mainly from the vmexit which is a hardware context switch between virtual machine and hypervisor. Our solution is to poll for a while and do not enter real idle path if we can get the schedule event during polling. Poll may cause the CPU waste so we adopt a smart polling mechanism to reduce the useless poll. Signed-off-by: Yang Zhang Signed-off-by: Quan Xu Cc: Juergen Gross Cc: Alok Kataria Cc: Rusty Russell Cc: Thomas Gleixner Cc: Ingo Molnar Cc: "H. Peter Anvin" Cc: x...@kernel.org Cc: virtualizat...@lists.linux-foundation.org Cc: linux-kernel@vger.kernel.org Cc: xen-de...@lists.xenproject.org Hmm, is the idle entry path really so critical to performance that a new pvops function is necessary? Juergen, Here is the data we get when running benchmark netperf: 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): 29031.6 bit/s -- 76.1 %CPU 2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0): 35787.7 bit/s -- 129.4 %CPU 3. w/ kvm dynamic poll: 35735.6 bit/s -- 200.0 %CPU 4. w/patch and w/ kvm dynamic poll: 42225.3 bit/s -- 198.7 %CPU 5. idle=poll 37081.7 bit/s -- 998.1 %CPU w/ this patch, we will improve performance by 23%.. even we could improve performance by 45.4%, if we use w/patch and w/ kvm dynamic poll. also the cost of CPU is much lower than 'idle=poll' case.. I don't question the general idea. I just think pvops isn't the best way to implement it. Wouldn't a function pointer, maybe guarded by a static key, be enough? A further advantage would be that this would work on other architectures, too. I assume this feature will be ported to other archs.. a new pvops makes sorry, a typo.. /other archs/other hypervisors/ it refers hypervisor like Xen, HyperV and VMware).. code clean and easy to maintain. also I tried to add it into existed pvops, but it doesn't match. You are aware that pvops is x86 only? yes, I'm aware.. I really don't see the big difference in maintainability compared to the static key / function pointer variant: void (*guest_idle_poll_func)(void); struct static_key guest_idle_poll_key __read_mostly; static inline void guest_idle_poll(void) { if (static_key_false(&guest_idle_poll_key)) guest_idle_poll_func(); } thank you for your sample code :) I agree there is no big difference.. I think we are discussion for two things: 1) x86 VM on different hypervisors 2) different archs VM on kvm hypervisor What I want to do is x86 VM on different hypervisors, such as kvm / xen / hyperv .. And KVM would just need to set guest_idle_poll_func and enable the static key. Works on non-x86 architectures, too. .. referred to 'pv_mmu_ops', HyperV and Xen can implement their own functions for 'pv_mmu_ops'. I think it is the same to pv_idle_ops. with above explaination, do you still think I need to define the static key/function pointer variant? btw, any interest to port it to Xen HVM guest? :) Quan Alibaba Cloud
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
2017-11-14 16:15 GMT+08:00 Quan Xu : > > > On 2017/11/14 15:12, Wanpeng Li wrote: >> >> 2017-11-14 15:02 GMT+08:00 Quan Xu : >>> >>> >>> On 2017/11/13 18:53, Juergen Gross wrote: On 13/11/17 11:06, Quan Xu wrote: > > From: Quan Xu > > So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called > in idle path which will poll for a while before we enter the real idle > state. > > In virtualization, idle path includes several heavy operations > includes timer access(LAPIC timer or TSC deadline timer) which will > hurt performance especially for latency intensive workload like message > passing task. The cost is mainly from the vmexit which is a hardware > context switch between virtual machine and hypervisor. Our solution is > to poll for a while and do not enter real idle path if we can get the > schedule event during polling. > > Poll may cause the CPU waste so we adopt a smart polling mechanism to > reduce the useless poll. > > Signed-off-by: Yang Zhang > Signed-off-by: Quan Xu > Cc: Juergen Gross > Cc: Alok Kataria > Cc: Rusty Russell > Cc: Thomas Gleixner > Cc: Ingo Molnar > Cc: "H. Peter Anvin" > Cc: x...@kernel.org > Cc: virtualizat...@lists.linux-foundation.org > Cc: linux-kernel@vger.kernel.org > Cc: xen-de...@lists.xenproject.org Hmm, is the idle entry path really so critical to performance that a new pvops function is necessary? >>> >>> Juergen, Here is the data we get when running benchmark netperf: >>> 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): >>> 29031.6 bit/s -- 76.1 %CPU >>> >>> 2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0): >>> 35787.7 bit/s -- 129.4 %CPU >>> >>> 3. w/ kvm dynamic poll: >>> 35735.6 bit/s -- 200.0 %CPU >> >> Actually we can reduce the CPU utilization by sleeping a period of >> time as what has already been done in the poll logic of IO subsystem, >> then we can improve the algorithm in kvm instead of introduing another >> duplicate one in the kvm guest. > > We really appreciate upstream's kvm dynamic poll mechanism, which is > really helpful for a lot of scenario.. > > However, as description said, in virtualization, idle path includes > several heavy operations includes timer access (LAPIC timer or TSC > deadline timer) which will hurt performance especially for latency > intensive workload like message passing task. The cost is mainly from > the vmexit which is a hardware context switch between virtual machine > and hypervisor. > > for upstream's kvm dynamic poll mechanism, even you could provide a > better algorism, how could you bypass timer access (LAPIC timer or TSC > deadline timer), or a hardware context switch between virtual machine > and hypervisor. I know these is a tradeoff. > > Furthermore, here is the data we get when running benchmark contextswitch > to measure the latency(lower is better): > > 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): > 3402.9 ns/ctxsw -- 199.8 %CPU > > 2. w/ patch and disable kvm dynamic poll: > 1163.5 ns/ctxsw -- 205.5 %CPU > > 3. w/ kvm dynamic poll: > 2280.6 ns/ctxsw -- 199.5 %CPU > > so, these tow solution are quite similar, but not duplicate.. > > that's also why to add a generic idle poll before enter real idle path. > When a reschedule event is pending, we can bypass the real idle path. > There is a similar logic in the idle governor/driver, so how this patchset influence the decision in the idle governor/driver when running on bare-metal(power managment is not exposed to the guest so we will not enter into idle driver in the guest)? Regards, Wanpeng Li
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
On 2017/11/14 15:12, Wanpeng Li wrote: 2017-11-14 15:02 GMT+08:00 Quan Xu : On 2017/11/13 18:53, Juergen Gross wrote: On 13/11/17 11:06, Quan Xu wrote: From: Quan Xu So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called in idle path which will poll for a while before we enter the real idle state. In virtualization, idle path includes several heavy operations includes timer access(LAPIC timer or TSC deadline timer) which will hurt performance especially for latency intensive workload like message passing task. The cost is mainly from the vmexit which is a hardware context switch between virtual machine and hypervisor. Our solution is to poll for a while and do not enter real idle path if we can get the schedule event during polling. Poll may cause the CPU waste so we adopt a smart polling mechanism to reduce the useless poll. Signed-off-by: Yang Zhang Signed-off-by: Quan Xu Cc: Juergen Gross Cc: Alok Kataria Cc: Rusty Russell Cc: Thomas Gleixner Cc: Ingo Molnar Cc: "H. Peter Anvin" Cc: x...@kernel.org Cc: virtualizat...@lists.linux-foundation.org Cc: linux-kernel@vger.kernel.org Cc: xen-de...@lists.xenproject.org Hmm, is the idle entry path really so critical to performance that a new pvops function is necessary? Juergen, Here is the data we get when running benchmark netperf: 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): 29031.6 bit/s -- 76.1 %CPU 2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0): 35787.7 bit/s -- 129.4 %CPU 3. w/ kvm dynamic poll: 35735.6 bit/s -- 200.0 %CPU Actually we can reduce the CPU utilization by sleeping a period of time as what has already been done in the poll logic of IO subsystem, then we can improve the algorithm in kvm instead of introduing another duplicate one in the kvm guest. We really appreciate upstream's kvm dynamic poll mechanism, which is really helpful for a lot of scenario.. However, as description said, in virtualization, idle path includes several heavy operations includes timer access (LAPIC timer or TSC deadline timer) which will hurt performance especially for latency intensive workload like message passing task. The cost is mainly from the vmexit which is a hardware context switch between virtual machine and hypervisor. for upstream's kvm dynamic poll mechanism, even you could provide a better algorism, how could you bypass timer access (LAPIC timer or TSC deadline timer), or a hardware context switch between virtual machine and hypervisor. I know these is a tradeoff. Furthermore, here is the data we get when running benchmark contextswitch to measure the latency(lower is better): 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): 3402.9 ns/ctxsw -- 199.8 %CPU 2. w/ patch and disable kvm dynamic poll: 1163.5 ns/ctxsw -- 205.5 %CPU 3. w/ kvm dynamic poll: 2280.6 ns/ctxsw -- 199.5 %CPU so, these tow solution are quite similar, but not duplicate.. that's also why to add a generic idle poll before enter real idle path. When a reschedule event is pending, we can bypass the real idle path. Quan Alibaba Cloud Regards, Wanpeng Li 4. w/patch and w/ kvm dynamic poll: 42225.3 bit/s -- 198.7 %CPU 5. idle=poll 37081.7 bit/s -- 998.1 %CPU w/ this patch, we will improve performance by 23%.. even we could improve performance by 45.4%, if we use w/patch and w/ kvm dynamic poll. also the cost of CPU is much lower than 'idle=poll' case.. Wouldn't a function pointer, maybe guarded by a static key, be enough? A further advantage would be that this would work on other architectures, too. I assume this feature will be ported to other archs.. a new pvops makes code clean and easy to maintain. also I tried to add it into existed pvops, but it doesn't match. Quan Alibaba Cloud Juergen
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
On 14/11/17 08:02, Quan Xu wrote: > > > On 2017/11/13 18:53, Juergen Gross wrote: >> On 13/11/17 11:06, Quan Xu wrote: >>> From: Quan Xu >>> >>> So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called >>> in idle path which will poll for a while before we enter the real idle >>> state. >>> >>> In virtualization, idle path includes several heavy operations >>> includes timer access(LAPIC timer or TSC deadline timer) which will >>> hurt performance especially for latency intensive workload like message >>> passing task. The cost is mainly from the vmexit which is a hardware >>> context switch between virtual machine and hypervisor. Our solution is >>> to poll for a while and do not enter real idle path if we can get the >>> schedule event during polling. >>> >>> Poll may cause the CPU waste so we adopt a smart polling mechanism to >>> reduce the useless poll. >>> >>> Signed-off-by: Yang Zhang >>> Signed-off-by: Quan Xu >>> Cc: Juergen Gross >>> Cc: Alok Kataria >>> Cc: Rusty Russell >>> Cc: Thomas Gleixner >>> Cc: Ingo Molnar >>> Cc: "H. Peter Anvin" >>> Cc: x...@kernel.org >>> Cc: virtualizat...@lists.linux-foundation.org >>> Cc: linux-kernel@vger.kernel.org >>> Cc: xen-de...@lists.xenproject.org >> Hmm, is the idle entry path really so critical to performance that a new >> pvops function is necessary? > Juergen, Here is the data we get when running benchmark netperf: > 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): > 29031.6 bit/s -- 76.1 %CPU > > 2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0): > 35787.7 bit/s -- 129.4 %CPU > > 3. w/ kvm dynamic poll: > 35735.6 bit/s -- 200.0 %CPU > > 4. w/patch and w/ kvm dynamic poll: > 42225.3 bit/s -- 198.7 %CPU > > 5. idle=poll > 37081.7 bit/s -- 998.1 %CPU > > > > w/ this patch, we will improve performance by 23%.. even we could improve > performance by 45.4%, if we use w/patch and w/ kvm dynamic poll. also the > cost of CPU is much lower than 'idle=poll' case.. I don't question the general idea. I just think pvops isn't the best way to implement it. >> Wouldn't a function pointer, maybe guarded >> by a static key, be enough? A further advantage would be that this would >> work on other architectures, too. > > I assume this feature will be ported to other archs.. a new pvops makes > code > clean and easy to maintain. also I tried to add it into existed pvops, > but it > doesn't match. You are aware that pvops is x86 only? I really don't see the big difference in maintainability compared to the static key / function pointer variant: void (*guest_idle_poll_func)(void); struct static_key guest_idle_poll_key __read_mostly; static inline void guest_idle_poll(void) { if (static_key_false(&guest_idle_poll_key)) guest_idle_poll_func(); } And KVM would just need to set guest_idle_poll_func and enable the static key. Works on non-x86 architectures, too. Juergen
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
2017-11-14 15:02 GMT+08:00 Quan Xu : > > > On 2017/11/13 18:53, Juergen Gross wrote: >> >> On 13/11/17 11:06, Quan Xu wrote: >>> >>> From: Quan Xu >>> >>> So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called >>> in idle path which will poll for a while before we enter the real idle >>> state. >>> >>> In virtualization, idle path includes several heavy operations >>> includes timer access(LAPIC timer or TSC deadline timer) which will >>> hurt performance especially for latency intensive workload like message >>> passing task. The cost is mainly from the vmexit which is a hardware >>> context switch between virtual machine and hypervisor. Our solution is >>> to poll for a while and do not enter real idle path if we can get the >>> schedule event during polling. >>> >>> Poll may cause the CPU waste so we adopt a smart polling mechanism to >>> reduce the useless poll. >>> >>> Signed-off-by: Yang Zhang >>> Signed-off-by: Quan Xu >>> Cc: Juergen Gross >>> Cc: Alok Kataria >>> Cc: Rusty Russell >>> Cc: Thomas Gleixner >>> Cc: Ingo Molnar >>> Cc: "H. Peter Anvin" >>> Cc: x...@kernel.org >>> Cc: virtualizat...@lists.linux-foundation.org >>> Cc: linux-kernel@vger.kernel.org >>> Cc: xen-de...@lists.xenproject.org >> >> Hmm, is the idle entry path really so critical to performance that a new >> pvops function is necessary? > > Juergen, Here is the data we get when running benchmark netperf: > 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): > 29031.6 bit/s -- 76.1 %CPU > > 2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0): > 35787.7 bit/s -- 129.4 %CPU > > 3. w/ kvm dynamic poll: > 35735.6 bit/s -- 200.0 %CPU Actually we can reduce the CPU utilization by sleeping a period of time as what has already been done in the poll logic of IO subsystem, then we can improve the algorithm in kvm instead of introduing another duplicate one in the kvm guest. Regards, Wanpeng Li > > 4. w/patch and w/ kvm dynamic poll: > 42225.3 bit/s -- 198.7 %CPU > > 5. idle=poll > 37081.7 bit/s -- 998.1 %CPU > > > > w/ this patch, we will improve performance by 23%.. even we could improve > performance by 45.4%, if we use w/patch and w/ kvm dynamic poll. also the > cost of CPU is much lower than 'idle=poll' case.. > >> Wouldn't a function pointer, maybe guarded >> by a static key, be enough? A further advantage would be that this would >> work on other architectures, too. > > > I assume this feature will be ported to other archs.. a new pvops makes code > clean and easy to maintain. also I tried to add it into existed pvops, but > it > doesn't match. > > > > Quan > Alibaba Cloud >> >> >> Juergen >> >
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
On 2017/11/13 18:53, Juergen Gross wrote: On 13/11/17 11:06, Quan Xu wrote: From: Quan Xu So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called in idle path which will poll for a while before we enter the real idle state. In virtualization, idle path includes several heavy operations includes timer access(LAPIC timer or TSC deadline timer) which will hurt performance especially for latency intensive workload like message passing task. The cost is mainly from the vmexit which is a hardware context switch between virtual machine and hypervisor. Our solution is to poll for a while and do not enter real idle path if we can get the schedule event during polling. Poll may cause the CPU waste so we adopt a smart polling mechanism to reduce the useless poll. Signed-off-by: Yang Zhang Signed-off-by: Quan Xu Cc: Juergen Gross Cc: Alok Kataria Cc: Rusty Russell Cc: Thomas Gleixner Cc: Ingo Molnar Cc: "H. Peter Anvin" Cc: x...@kernel.org Cc: virtualizat...@lists.linux-foundation.org Cc: linux-kernel@vger.kernel.org Cc: xen-de...@lists.xenproject.org Hmm, is the idle entry path really so critical to performance that a new pvops function is necessary? Juergen, Here is the data we get when running benchmark netperf: 1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0): 29031.6 bit/s -- 76.1 %CPU 2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0): 35787.7 bit/s -- 129.4 %CPU 3. w/ kvm dynamic poll: 35735.6 bit/s -- 200.0 %CPU 4. w/patch and w/ kvm dynamic poll: 42225.3 bit/s -- 198.7 %CPU 5. idle=poll 37081.7 bit/s -- 998.1 %CPU w/ this patch, we will improve performance by 23%.. even we could improve performance by 45.4%, if we use w/patch and w/ kvm dynamic poll. also the cost of CPU is much lower than 'idle=poll' case.. Wouldn't a function pointer, maybe guarded by a static key, be enough? A further advantage would be that this would work on other architectures, too. I assume this feature will be ported to other archs.. a new pvops makes code clean and easy to maintain. also I tried to add it into existed pvops, but it doesn't match. Quan Alibaba Cloud Juergen
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
2017-11-13 18:53 GMT+08:00 Juergen Gross : > On 13/11/17 11:06, Quan Xu wrote: >> From: Quan Xu >> >> So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called >> in idle path which will poll for a while before we enter the real idle >> state. >> >> In virtualization, idle path includes several heavy operations >> includes timer access(LAPIC timer or TSC deadline timer) which will >> hurt performance especially for latency intensive workload like message >> passing task. The cost is mainly from the vmexit which is a hardware >> context switch between virtual machine and hypervisor. Our solution is >> to poll for a while and do not enter real idle path if we can get the >> schedule event during polling. >> >> Poll may cause the CPU waste so we adopt a smart polling mechanism to >> reduce the useless poll. >> >> Signed-off-by: Yang Zhang >> Signed-off-by: Quan Xu >> Cc: Juergen Gross >> Cc: Alok Kataria >> Cc: Rusty Russell >> Cc: Thomas Gleixner >> Cc: Ingo Molnar >> Cc: "H. Peter Anvin" >> Cc: x...@kernel.org >> Cc: virtualizat...@lists.linux-foundation.org >> Cc: linux-kernel@vger.kernel.org >> Cc: xen-de...@lists.xenproject.org > > Hmm, is the idle entry path really so critical to performance that a new > pvops function is necessary? Wouldn't a function pointer, maybe guarded > by a static key, be enough? A further advantage would be that this would > work on other architectures, too. There is a "Adaptive halt-polling" which are merged to upstream more than two years ago avoids to thread the critical path and has already been ported to other architectures. https://lkml.org/lkml/2015/9/3/615 Regards, Wanpeng Li
Re: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
On 13/11/17 11:06, Quan Xu wrote: > From: Quan Xu > > So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called > in idle path which will poll for a while before we enter the real idle > state. > > In virtualization, idle path includes several heavy operations > includes timer access(LAPIC timer or TSC deadline timer) which will > hurt performance especially for latency intensive workload like message > passing task. The cost is mainly from the vmexit which is a hardware > context switch between virtual machine and hypervisor. Our solution is > to poll for a while and do not enter real idle path if we can get the > schedule event during polling. > > Poll may cause the CPU waste so we adopt a smart polling mechanism to > reduce the useless poll. > > Signed-off-by: Yang Zhang > Signed-off-by: Quan Xu > Cc: Juergen Gross > Cc: Alok Kataria > Cc: Rusty Russell > Cc: Thomas Gleixner > Cc: Ingo Molnar > Cc: "H. Peter Anvin" > Cc: x...@kernel.org > Cc: virtualizat...@lists.linux-foundation.org > Cc: linux-kernel@vger.kernel.org > Cc: xen-de...@lists.xenproject.org Hmm, is the idle entry path really so critical to performance that a new pvops function is necessary? Wouldn't a function pointer, maybe guarded by a static key, be enough? A further advantage would be that this would work on other architectures, too. Juergen