| Message ID | 20210831015919.13006-1-skyele@sjtu.edu.cn (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/4] KVM: x86: Introduce .pcpu_is_idle() stub infrastructure | expand |
On Tue, 2021-08-31 at 09:59 +0800, Tianqiang Xu wrote: > This patch series aims to fix performance issue caused by current > para-virtualized scheduling design. Series? This looks to me like a patch, not a cover letter. If you want a cover letter, please make one. /Jarkko
Hi Tianqiang, Tianqiang Xu <skyele@sjtu.edu.cn> 于2021年12月17日周五 15:55写道： > > This patch series aims to fix performance issue caused by current > para-virtualized scheduling design. > > The current para-virtualized scheduling design uses 'preempted' field of > kvm_steal_time to avoid scheduling task on the preempted vCPU. > However, when the pCPU where the preempted vCPU most recently run is idle, > it will result in low cpu utilization, and consequently poor performance. > > The new field: 'is_idle' of kvm_steal_time can precisely reveal > the status of pCPU where preempted vCPU most recently run, and > then improve cpu utilization. > > pcpu_is_idle() is used to get the value of 'is_idle' of kvm_steal_time. > > Experiments on a VM with 16 vCPUs show that the patch can reduce around > 50% to 80% execution time for most PARSEC benchmarks. > This also holds true for a VM with 112 vCPUs. > > Experiments on 2 VMs with 112 vCPUs show that the patch can reduce around > 20% to 80% execution time for most PARSEC benchmarks. > > Test environment: > -- PowerEdge R740 > -- 56C-112T CPU Intel(R) Xeon(R) Gold 6238R CPU > -- Host 190G DRAM > -- QEMU 5.0.0 > -- PARSEC 3.0 Native Inputs > -- Host is idle during the test > -- Host and Guest kernel are both kernel-5.14.0 > > Results: > 1. 1 VM, 16 VCPU, 16 THREAD. > Host Topology: sockets=2 cores=28 threads=2 > VM Topology: sockets=1 cores=16 threads=1 > Command: <path to parsec>/bin/parsecmgmt -a run -p <benchmark> -i native -n 16 > Statistics below are the real time of running each benchmark.(lower is better) > > before patch after patch improvements > bodytrack 52.866s 22.619s 57.21% > fluidanimate 84.009s 38.148s 54.59% > streamcluster 270.17s 42.726s 84.19% > splash2x.ocean_cp 31.932s 9.539s 70.13% > splash2x.ocean_ncp 36.063s 14.189s 60.65% > splash2x.volrend 134.587s 21.79s 83.81% > > 2. 1VM, 112 VCPU. Some benchmarks require the number of threads to be the power of 2, > so we run them with 64 threads and 128 threads. > Host Topology: sockets=2 cores=28 threads=2 > VM Topology: sockets=1 cores=112 threads=1 > Command: <path to parsec>/bin/parsecmgmt -a run -p <benchmark> -i native -n <64,112,128> > Statistics below are the real time of running each benchmark.(lower is better) > > before patch after patch improvements > fluidanimate(64 thread) 124.235s 27.924s 77.52% > fluidanimate(128 thread) 169.127s 64.541s 61.84% > streamcluster(112 thread) 861.879s 496.66s 42.37% > splash2x.ocean_cp(64 thread) 46.415s 18.527s 60.08% > splash2x.ocean_cp(128 thread) 53.647s 28.929s 46.08% > splash2x.ocean_ncp(64 thread) 47.613s 19.576s 58.89% > splash2x.ocean_ncp(128 thread) 54.94s 29.199s 46.85% > splash2x.volrend(112 thread) 801.384s 144.824s 81.93% > > 3. 2VM, each VM: 112 VCPU. Some benchmarks require the number of threads to > be the power of 2, so we run them with 64 threads and 128 threads. > Host Topology: sockets=2 cores=28 threads=2 > VM Topology: sockets=1 cores=112 threads=1 > Command: <path to parsec>/bin/parsecmgmt -a run -p <benchmark> -i native -n <64,112,128> > Statistics below are the average real time of running each benchmark in 2 VMs.(lower is better) > > before patch after patch improvements > fluidanimate(64 thread) 135.2125s 49.827s 63.15% > fluidanimate(128 thread) 178.309s 86.964s 51.23% > splash2x.ocean_cp(64 thread) 47.4505s 20.314s 57.19% > splash2x.ocean_cp(128 thread) 55.5645s 30.6515s 44.84% > splash2x.ocean_ncp(64 thread) 49.9775s 23.489s 53.00% > splash2x.ocean_ncp(128 thread) 56.847s 28.545s 49.79% > splash2x.volrend(112 thread) 838.939s 239.632s 71.44% > > For space limit, we list representative statistics here. I did a performance test according to the description in the patch, but did not get the performance improvement described in the description. I suspect that the big difference between my kernel configuration and yours has caused this problem. Can you please provide more detailed test information, such as kernel configuration that must be turned on or off ? Regards, Jinrong Liang
diff --git a/arch/x86/hyperv/hv_spinlock.c b/arch/x86/hyperv/hv_spinlock.c index 91cfe698bde0..9782e188904a 100644 --- a/arch/x86/hyperv/hv_spinlock.c +++ b/arch/x86/hyperv/hv_spinlock.c @@ -60,6 +60,12 @@ static void hv_qlock_wait(u8 *byte, u8 val) /* * Hyper-V does not support this so far. */ +__visible bool hv_pcpu_is_idle(int vcpu) +{ + return false; +} +PV_CALLEE_SAVE_REGS_THUNK(hv_pcpu_is_idle); + __visible bool hv_vcpu_is_preempted(int vcpu) { return false; @@ -82,6 +88,7 @@ void __init hv_init_spinlocks(void) pv_ops.lock.wait = hv_qlock_wait; pv_ops.lock.kick = hv_qlock_kick; pv_ops.lock.vcpu_is_preempted = PV_CALLEE_SAVE(hv_vcpu_is_preempted); + pv_ops.lock.pcpu_is_idle = PV_CALLEE_SAVE(hv_pcpu_is_idle); } static __init int hv_parse_nopvspin(char *arg) diff --git a/arch/x86/include/asm/cpufeatures.h b/arch/x86/include/asm/cpufeatures.h index d0ce5cfd3ac1..efda9b9a4cad 100644 --- a/arch/x86/include/asm/cpufeatures.h +++ b/arch/x86/include/asm/cpufeatures.h @@ -238,6 +238,7 @@ #define X86_FEATURE_VMW_VMMCALL ( 8*32+19) /* "" VMware prefers VMMCALL hypercall instruction */ #define X86_FEATURE_PVUNLOCK ( 8*32+20) /* "" PV unlock function */ #define X86_FEATURE_VCPUPREEMPT ( 8*32+21) /* "" PV vcpu_is_preempted function */ +#define X86_FEATURE_PCPUISIDLE ( 8*32+22) /* "" PV pcpu_is_idle function */ /* Intel-defined CPU features, CPUID level 0x00000007:0 (EBX), word 9 */ #define X86_FEATURE_FSGSBASE ( 9*32+ 0) /* RDFSBASE, WRFSBASE, RDGSBASE, WRGSBASE instructions*/ diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h index af6ce8d4c86a..705c55be0eed 100644 --- a/arch/x86/include/asm/kvm_host.h +++ b/arch/x86/include/asm/kvm_host.h @@ -742,6 +742,7 @@ struct kvm_vcpu_arch { struct { u8 preempted; + u8 is_idle; u64 msr_val; u64 last_steal; struct gfn_to_pfn_cache cache; diff --git a/arch/x86/include/asm/paravirt.h b/arch/x86/include/asm/paravirt.h index da3a1ac82be5..f34dec6eb515 100644 --- a/arch/x86/include/asm/paravirt.h +++ b/arch/x86/include/asm/paravirt.h @@ -609,8 +609,16 @@ static __always_inline bool pv_vcpu_is_preempted(long cpu) ALT_NOT(X86_FEATURE_VCPUPREEMPT)); } +static __always_inline bool pv_pcpu_is_idle(long cpu) +{ + return PVOP_ALT_CALLEE1(bool, lock.pcpu_is_idle, cpu, + "xor %%" _ASM_AX ", %%" _ASM_AX ";", + ALT_NOT(X86_FEATURE_PCPUISIDLE)); +} + void __raw_callee_save___native_queued_spin_unlock(struct qspinlock *lock); bool __raw_callee_save___native_vcpu_is_preempted(long cpu); +bool __raw_callee_save___native_pcpu_is_idle(long cpu); #endif /* SMP && PARAVIRT_SPINLOCKS */ diff --git a/arch/x86/include/asm/paravirt_types.h b/arch/x86/include/asm/paravirt_types.h index d9d6b0203ec4..7d9b5906580c 100644 --- a/arch/x86/include/asm/paravirt_types.h +++ b/arch/x86/include/asm/paravirt_types.h @@ -257,6 +257,7 @@ struct pv_lock_ops { void (*kick)(int cpu); struct paravirt_callee_save vcpu_is_preempted; + struct paravirt_callee_save pcpu_is_idle; } __no_randomize_layout; /* This contains all the paravirt structures: we get a convenient diff --git a/arch/x86/include/asm/qspinlock.h b/arch/x86/include/asm/qspinlock.h index d86ab942219c..c32f2eb6186c 100644 --- a/arch/x86/include/asm/qspinlock.h +++ b/arch/x86/include/asm/qspinlock.h @@ -63,6 +63,13 @@ static inline bool vcpu_is_preempted(long cpu) } #endif +#define pcpu_is_idle pcpu_is_idle +static inline bool pcpu_is_idle(long cpu) +{ + return pv_pcpu_is_idle(cpu); +} +#endif + #ifdef CONFIG_PARAVIRT /* * virt_spin_lock_key - enables (by default) the virt_spin_lock() hijack. diff --git a/arch/x86/include/uapi/asm/kvm_para.h b/arch/x86/include/uapi/asm/kvm_para.h index 5146bbab84d4..82940e4b76d5 100644 --- a/arch/x86/include/uapi/asm/kvm_para.h +++ b/arch/x86/include/uapi/asm/kvm_para.h @@ -63,12 +63,14 @@ struct kvm_steal_time { __u32 version; __u32 flags; __u8 preempted; - __u8 u8_pad[3]; + __u8 is_idle; + __u8 u8_pad[2]; __u32 pad[11]; }; #define KVM_VCPU_PREEMPTED (1 << 0) #define KVM_VCPU_FLUSH_TLB (1 << 1) +#define KVM_PCPU_IS_IDLE (1 << 0) #define KVM_CLOCK_PAIRING_WALLCLOCK 0 struct kvm_clock_pairing { diff --git a/arch/x86/kernel/asm-offsets_64.c b/arch/x86/kernel/asm-offsets_64.c index b14533af7676..b587bbe44470 100644 --- a/arch/x86/kernel/asm-offsets_64.c +++ b/arch/x86/kernel/asm-offsets_64.c @@ -22,6 +22,7 @@ int main(void) #if defined(CONFIG_KVM_GUEST) && defined(CONFIG_PARAVIRT_SPINLOCKS) OFFSET(KVM_STEAL_TIME_preempted, kvm_steal_time, preempted); + OFFSET(KVM_STEAL_TIME_is_idle, kvm_steal_time, is_idle); BLANK(); #endif diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c index a26643dc6bd6..b167589fffbc 100644 --- a/arch/x86/kernel/kvm.c +++ b/arch/x86/kernel/kvm.c @@ -900,11 +900,19 @@ __visible bool __kvm_vcpu_is_preempted(long cpu) } PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted); +__visible bool __kvm_pcpu_is_idle(long cpu) +{ + struct kvm_steal_time *src = &per_cpu(steal_time, cpu); + + return !!(src->preempted & KVM_PCPU_IS_IDLE); +} +PV_CALLEE_SAVE_REGS_THUNK(__kvm_pcpu_is_idle); #else #include <asm/asm-offsets.h> extern bool __raw_callee_save___kvm_vcpu_is_preempted(long); +extern bool __raw_callee_save___kvm_pcpu_is_idle(long); /* * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and @@ -922,6 +930,17 @@ asm( ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;" ".popsection"); +asm( +".pushsection .text;" +".global __raw_callee_save___kvm_pcpu_is_idle;" +".type __raw_callee_save___kvm_pcpu_is_idle, @function;" +"__raw_callee_save___kvm_pcpu_is_idle:" +"movq __per_cpu_offset(,%rdi,8), %rax;" +"cmpb $0, " __stringify(KVM_STEAL_TIME_is_idle) "+steal_time(%rax);" +"setne %al;" +"ret;" +".size __raw_callee_save___kvm_pcpu_is_idle, .-__raw_callee_save___kvm_pcpu_is_idle;" +".popsection"); #endif /* @@ -970,6 +989,8 @@ void __init kvm_spinlock_init(void) if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) { pv_ops.lock.vcpu_is_preempted = PV_CALLEE_SAVE(__kvm_vcpu_is_preempted); + pv_ops.lock.pcpu_is_idle = + PV_CALLEE_SAVE(__kvm_pcpu_is_idle); } /* * When PV spinlock is enabled which is preferred over diff --git a/arch/x86/kernel/paravirt-spinlocks.c b/arch/x86/kernel/paravirt-spinlocks.c index 9e1ea99ad9df..d7f6a461d0a5 100644 --- a/arch/x86/kernel/paravirt-spinlocks.c +++ b/arch/x86/kernel/paravirt-spinlocks.c @@ -27,12 +27,24 @@ __visible bool __native_vcpu_is_preempted(long cpu) } PV_CALLEE_SAVE_REGS_THUNK(__native_vcpu_is_preempted); +__visible bool __native_pcpu_is_idle(long cpu) +{ + return false; +} +PV_CALLEE_SAVE_REGS_THUNK(__native_pcpu_is_idle); + bool pv_is_native_vcpu_is_preempted(void) { return pv_ops.lock.vcpu_is_preempted.func == __raw_callee_save___native_vcpu_is_preempted; } +bool pv_is_native_pcpu_is_idle(void) +{ + return pv_ops.lock.pcpu_is_idle.func == + __raw_callee_save___native_pcpu_is_idle; +} + void __init paravirt_set_cap(void) { if (!pv_is_native_spin_unlock()) @@ -40,4 +52,7 @@ void __init paravirt_set_cap(void) if (!pv_is_native_vcpu_is_preempted()) setup_force_cpu_cap(X86_FEATURE_VCPUPREEMPT); + + if (!pv_is_native_pcpu_is_idle()) + setup_force_cpu_cap(X86_FEATURE_PCPUISIDLE); } diff --git a/arch/x86/kernel/paravirt.c b/arch/x86/kernel/paravirt.c index 04cafc057bed..543d89856161 100644 --- a/arch/x86/kernel/paravirt.c +++ b/arch/x86/kernel/paravirt.c @@ -366,6 +366,8 @@ struct paravirt_patch_template pv_ops = { .lock.kick = paravirt_nop, .lock.vcpu_is_preempted = PV_CALLEE_SAVE(__native_vcpu_is_preempted), + .lock.pcpu_is_idle = + PV_CALLEE_SAVE(__native_pcpu_is_idle), #endif /* SMP */ #endif };
This patch series aims to fix performance issue caused by current para-virtualized scheduling design. The current para-virtualized scheduling design uses 'preempted' field of kvm_steal_time to avoid scheduling task on the preempted vCPU. However, when the pCPU where the preempted vCPU most recently run is idle, it will result in low cpu utilization, and consequently poor performance. The new field: 'is_idle' of kvm_steal_time can precisely reveal the status of pCPU where preempted vCPU most recently run, and then improve cpu utilization. pcpu_is_idle() is used to get the value of 'is_idle' of kvm_steal_time. Experiments on a VM with 16 vCPUs show that the patch can reduce around 50% to 80% execution time for most PARSEC benchmarks. This also holds true for a VM with 112 vCPUs. Experiments on 2 VMs with 112 vCPUs show that the patch can reduce around 20% to 80% execution time for most PARSEC benchmarks. Test environment: -- PowerEdge R740 -- 56C-112T CPU Intel(R) Xeon(R) Gold 6238R CPU -- Host 190G DRAM -- QEMU 5.0.0 -- PARSEC 3.0 Native Inputs -- Host is idle during the test -- Host and Guest kernel are both kernel-5.14.0 Results: 1. 1 VM, 16 VCPU, 16 THREAD. Host Topology: sockets=2 cores=28 threads=2 VM Topology: sockets=1 cores=16 threads=1 Command: <path to parsec>/bin/parsecmgmt -a run -p <benchmark> -i native -n 16 Statistics below are the real time of running each benchmark.(lower is better) before patch after patch improvements bodytrack 52.866s 22.619s 57.21% fluidanimate 84.009s 38.148s 54.59% streamcluster 270.17s 42.726s 84.19% splash2x.ocean_cp 31.932s 9.539s 70.13% splash2x.ocean_ncp 36.063s 14.189s 60.65% splash2x.volrend 134.587s 21.79s 83.81% 2. 1VM, 112 VCPU. Some benchmarks require the number of threads to be the power of 2, so we run them with 64 threads and 128 threads. Host Topology: sockets=2 cores=28 threads=2 VM Topology: sockets=1 cores=112 threads=1 Command: <path to parsec>/bin/parsecmgmt -a run -p <benchmark> -i native -n <64,112,128> Statistics below are the real time of running each benchmark.(lower is better) before patch after patch improvements fluidanimate(64 thread) 124.235s 27.924s 77.52% fluidanimate(128 thread) 169.127s 64.541s 61.84% streamcluster(112 thread) 861.879s 496.66s 42.37% splash2x.ocean_cp(64 thread) 46.415s 18.527s 60.08% splash2x.ocean_cp(128 thread) 53.647s 28.929s 46.08% splash2x.ocean_ncp(64 thread) 47.613s 19.576s 58.89% splash2x.ocean_ncp(128 thread) 54.94s 29.199s 46.85% splash2x.volrend(112 thread) 801.384s 144.824s 81.93% 3. 2VM, each VM: 112 VCPU. Some benchmarks require the number of threads to be the power of 2, so we run them with 64 threads and 128 threads. Host Topology: sockets=2 cores=28 threads=2 VM Topology: sockets=1 cores=112 threads=1 Command: <path to parsec>/bin/parsecmgmt -a run -p <benchmark> -i native -n <64,112,128> Statistics below are the average real time of running each benchmark in 2 VMs.(lower is better) before patch after patch improvements fluidanimate(64 thread) 135.2125s 49.827s 63.15% fluidanimate(128 thread) 178.309s 86.964s 51.23% splash2x.ocean_cp(64 thread) 47.4505s 20.314s 57.19% splash2x.ocean_cp(128 thread) 55.5645s 30.6515s 44.84% splash2x.ocean_ncp(64 thread) 49.9775s 23.489s 53.00% splash2x.ocean_ncp(128 thread) 56.847s 28.545s 49.79% splash2x.volrend(112 thread) 838.939s 239.632s 71.44% For space limit, we list representative statistics here. -- Authors: Tianqiang Xu, Dingji Li, Zeyu Mi Shanghai Jiao Tong University Signed-off-by: Tianqiang Xu <skyele@sjtu.edu.cn> --- arch/x86/hyperv/hv_spinlock.c | 7 +++++++ arch/x86/include/asm/cpufeatures.h | 1 + arch/x86/include/asm/kvm_host.h | 1 + arch/x86/include/asm/paravirt.h | 8 ++++++++ arch/x86/include/asm/paravirt_types.h | 1 + arch/x86/include/asm/qspinlock.h | 7 +++++++ arch/x86/include/uapi/asm/kvm_para.h | 4 +++- arch/x86/kernel/asm-offsets_64.c | 1 + arch/x86/kernel/kvm.c | 21 +++++++++++++++++++++ arch/x86/kernel/paravirt-spinlocks.c | 15 +++++++++++++++ arch/x86/kernel/paravirt.c | 2 ++ 11 files changed, 67 insertions(+), 1 deletion(-)