x86/fpu: track AVX-512 usage of tasks
Subsystems (3)
PATCH v6 1/3 x86/fpu: track AVX-512 usage of tasks
User space tools which do automated task placement need information about AVX-512 usage of tasks, because AVX-512 usage could cause core turbo frequency drop and impact the running task on the sibling CPU.
The XSAVE hardware structure has bits that indicate when valid state is present in registers unique to AVX-512 use. Use these bits to indicate when AVX-512 has been in use and add per-task AVX-512 state timestamp tracking to context switch.
Well-written AVX-512 applications are expected to clear the AVX-512 state when not actively using AVX-512 registers, so the tracking mechanism is imprecise and can theoretically miss AVX-512 usage during context switch. But it has been measured to be precise enough to be useful under real-world workloads like tensorflow and linpack.
If higher precision is required, suggest user space tools to use the PMU-based mechanisms in combination.
--- arch/x86/include/asm/fpu/internal.h | 7 +++++++ arch/x86/include/asm/fpu/types.h | 7 +++++++ 2 files changed, 14 insertions(+)
arch/x86/include/asm/fpu/internal.h +7 -0
@@ -411,6 +411,13 @@ static inline int copy_fpregs_to_fpstate(struct fpu *fpu) {if (likely (use_xsave ())) {copy_xregs_to_kernel (&fpu->state.xsave); + +/* +* AVX512 state is tracked here because its use is +* known to slow the max clock speed of the core. +*/ +if (fpu->state.xsave.header.xfeatures &XFEATURE_MASK_AVX512 ) + fpu->avx512_timestamp = jiffies_64;return 1 ; }
arch/x86/include/asm/fpu/types.h +7 -0
@@ -302,6 +302,13 @@ struct fpu {*/ unsigned char initialized; +/* +* @avx512_timestamp: +* +* Records the timestamp of AVX512 use during last context switch. +*/ +u64 avx512_timestamp; +/* * @state: *
Signature · 1 line
2.17.1
PATCH v6 2/3 proc: add AVX-512 usage elapsed time to /proc/pid/status
AVX-512 components use could cause core turbo frequency drop. So it's useful to expose AVX-512 usage elapsed time as a heuristic hint for the user space job scheduler to cluster the AVX-512 using tasks together.
Example:
$ cat /proc/pid/status | grep AVX512_elapsed_ms AVX512_elapsed_ms: 1020
The number '1020' denotes 1020 millisecond elapsed since last time context switch the off-CPU task using AVX-512 components, thus the task could cause core frequency drop.
Or:
$ cat /proc/pid/status | grep AVX512_elapsed_ms AVX512_elapsed_ms: -1
The number '-1' indicates the task didn't use AVX-512 components before thus unlikely has frequency drop issue.
User space tools may want to further check by:
$ perf stat --pid <pid> -e core_power.lvl2_turbo_license -- sleep 1
Performance counter stats for process id '3558':
3,251,565,961 core_power.lvl2_turbo_license
1.004031387 seconds time elapsedNon-zero counter value confirms that the task causes frequency drop.
--- arch/x86/kernel/fpu/xstate.c | 34 ++++++++++++++++++++++++++++++++++ fs/proc/array.c | 5 +++++ 2 files changed, 39 insertions(+)
arch/x86/kernel/fpu/xstate.c +34 -0
@@ -7,6 +7,7 @@#include <linux/cpu.h> #include <linux/mman.h> #include <linux/pkeys.h> +#include <linux/seq_file.h> #include <asm/fpu/api.h> #include <asm/fpu/internal.h>
@@ -1245,3 +1246,36 @@ int copy_user_to_xstate(struct xregs_state *xsave, const void __user *ubuf)return 0 ; } + +/* +* Report the amount of time elapsed in millisecond since last AVX512 +* use in the task. +*/ +void avx512_state (struct seq_file *m,struct task_struct *task) +{ +u64 timestamp = task->thread.fpu.avx512_timestamp; +s64 delta; + +if (!timestamp) + delta =-1 ; +else { +WARN_ON_ONCE (jiffies_64 < timestamp); + delta =div_u64 (jiffies64_to_nsecs (jiffies_64 - timestamp), +NSEC_PER_MSEC ); + } + +seq_put_decimal_ll (m,"AVX512_elapsed_ms:\t" , delta); +seq_putc (m,'\n' ); +} + +/* +* Report CPU specific thread state +*/ +void arch_task_state (struct seq_file *m,struct task_struct *task) +{ +/* +* Report AVX512 state if the processor and build option supported. +*/ +if (cpu_feature_enabled (X86_FEATURE_AVX512F )) +avx512_state (m, task); +}
fs/proc/array.c +5 -0
@@ -392,6 +392,10 @@ static inline void task_core_dumping(struct seq_file *m, struct mm_struct *mm)seq_putc (m,'\n' ); } +void __weak arch_task_state (struct seq_file *m,struct task_struct *task) +{ +} +int proc_pid_status (struct seq_file *m,struct pid_namespace *ns,struct pid *pid,struct task_struct *task) {
@@ -414,6 +418,7 @@ int proc_pid_status(struct seq_file *m, struct pid_namespace *ns,task_cpus_allowed (m, task);cpuset_task_status_allowed (m, task);task_context_switch_counts (m, task); +arch_task_state (m, task);return 0 ; }
Signature · 1 line
2.17.1
PATCH v6 3/3 Documentation/filesystems/proc.txt: add AVX512_elapsed_ms
In reply to Aubrey LiAdded AVX512_elapsed_ms in /proc/<pid>/status. Report it in Documentation/filesystems/proc.txt
--- Documentation/filesystems/proc.txt | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-)
Documentation/filesystems/proc.txt +3 -1
@@ -197,6 +197,7 @@ read the file /proc/PID/status: Seccomp: 0 voluntary_ctxt_switches: 0 nonvoluntary_ctxt_switches: 1 + AVX512_elapsed_ms: 1020 This shows you nearly the same information you would get if you viewed it with the ps command. In fact, ps uses the proc file system to obtain its
@@ -214,7 +215,7 @@ asynchronous manner and the value may not be very precise. To see a precise snapshot of a moment, you can see /proc/<pid>/smaps file and scan page table. It's slow but very precise. -Table 1-2: Contents of the status files (as of 4.8) +Table 1-2: Contents of the status files (as of 4.21) .............................................................................. Field Content Name filename of the executable
@@ -275,6 +276,7 @@ Table 1-2: Contents of the status files (as of 4.8) Mems_allowed_list Same as previous, but in "list format" voluntary_ctxt_switches number of voluntary context switches nonvoluntary_ctxt_switches number of non voluntary context switches + AVX512_elapsed_ms time elapsed since last AVX512 use in millisecond .............................................................................. Table 1-3: Contents of the statm files (as of 2.6.8-rc3)
Signature · 1 line
2.17.1
Quoted 5 lines · Aubrey Li
arch/x86/include/asm/fpu/internal.h +0
-0
arch/x86/include/asm/fpu/internal.h static inline int copy_fpregs_to_fpstate(struct fpu *fpu) · line 420
Quoted 10 lines
arch/x86/include/asm/fpu/internal.h +7
-0@@ -411,6 +411,13 @@ static inline int copy_fpregs_to_fpstate(struct fpu *fpu) {if (likely (use_xsave ())) {copy_xregs_to_kernel (&fpu->state.xsave); + +/* +* AVX512 state is tracked here because its use is +* known to slow the max clock speed of the core. +*/ +if (fpu->state.xsave.header.xfeatures &XFEATURE_MASK_AVX512 ) + fpu->avx512_timestamp = jiffies_64;
Even if unlikely this is incorrect when running a 32 bit kernel because there jiffies_64 cannot be atomically loaded vs. a concurrent update. See the comment in include/linux/jiffies.h right above the jiffies_64 declaration.
Thanks,
tglx
Quoted 13 lines · Thomas Gleixner
Quoted 12 lines · Aubrey Li
arch/x86/include/asm/fpu/internal.h +4
-0@@ -411,6 +411,13 @@ static inline int copy_fpregs_to_fpstate(struct fpu *fpu) {if (likely (use_xsave ())) {copy_xregs_to_kernel (&fpu->state.xsave); + +/* +* AVX512 state is tracked here because its use is +* known to slow the max clock speed of the core.
Quoted 9 lines
arch/x86/include/asm/fpu/internal.h static inline int copy_fpregs_to_fpstate(struct fpu *fpu) · line 420
Quoted 3 lines
arch/x86/include/asm/fpu/internal.h +3
-0+*/ +if (fpu->state.xsave.header.xfeatures &XFEATURE_MASK_AVX512 ) + fpu->avx512_timestamp = jiffies_64;Even if unlikely this is incorrect when running a 32 bit kernel because there
jiffies_64cannot be atomically loaded vs. a concurrent update. See the comment ininclude/linux/jiffies.hright above thejiffies_64declaration.
Yeah, I noticed this, because this is under use_xsave() condition, also need valid AVX512 state, so a 32 bit kernel won't enter this branch.
Thanks,
-Aubrey
Quoted 23 lines · Aubrey
Quoted 13 lines · Thomas Gleixner
Quoted 12 lines · Aubrey Li
arch/x86/include/asm/fpu/internal.h +4
-0@@ -411,6 +411,13 @@ static inline int copy_fpregs_to_fpstate(struct fpu *fpu) {if (likely (use_xsave ())) {copy_xregs_to_kernel (&fpu->state.xsave); + +/* +* AVX512 state is tracked here because its use is +* known to slow the max clock speed of the core. Quoted 9 lines
arch/x86/include/asm/fpu/internal.h static inline int copy_fpregs_to_fpstate(struct fpu *fpu) · line 420
Quoted 3 lines
arch/x86/include/asm/fpu/internal.h +3
-0+*/ +if (fpu->state.xsave.header.xfeatures &XFEATURE_MASK_AVX512 ) + fpu->avx512_timestamp = jiffies_64;Even if unlikely this is incorrect when running a 32 bit kernel because there
jiffies_64cannot be atomically loaded vs. a concurrent update. See the comment ininclude/linux/jiffies.hright above thejiffies_64declaration.
Quoted 2 lines
Yeah, I noticed this, because this is under
use_xsave()condition, also need valid AVX512 state, so a 32 bit kernel won't enter this branch.
What exactly prevents a 32bit kernel from having the AVX512 feature bit set? And if it cannot be set on 32bit, then why are you compiling that code in at all?
Thanks,
tglx
Quoted 25 lines · Thomas Gleixner
Quoted 23 lines · Aubrey
Quoted 22 lines · Thomas Gleixner
Quoted 5 lines · Aubrey Li
arch/x86/include/asm/fpu/internal.h +0
-0arch/x86/include/asm/fpu/internal.h static inline int copy_fpregs_to_fpstate(struct fpu *fpu) · line 420
Quoted 10 lines
arch/x86/include/asm/fpu/internal.h +7
-0@@ -411,6 +411,13 @@ static inline int copy_fpregs_to_fpstate(struct fpu *fpu) {if (likely (use_xsave ())) {copy_xregs_to_kernel (&fpu->state.xsave); + +/* +* AVX512 state is tracked here because its use is +* known to slow the max clock speed of the core. +*/ +if (fpu->state.xsave.header.xfeatures &XFEATURE_MASK_AVX512 ) + fpu->avx512_timestamp = jiffies_64;Even if unlikely this is incorrect when running a 32 bit kernel because there
jiffies_64cannot be atomically loaded vs. a concurrent update. See the comment ininclude/linux/jiffies.hright above thejiffies_64declaration.
Quoted 7 lines
Quoted 2 lines
Yeah, I noticed this, because this is under
use_xsave()condition, also need valid AVX512 state, so a 32 bit kernel won't enter this branch.What exactly prevents a 32bit kernel from having the AVX512 feature bit set? And if it cannot be set on 32bit, then why are you compiling that code in at all?
I misunderstood, you mean 32bit kernel, not 32bit machine. Theoretically 32bit kernel can use AVX512, but not sure if anyone use it like this. get_jiffies_64() includes jiffies_lock ops so not good in context switch. So I want to use raw jiffies_64 here. jiffies is a good candidate but it has wraparound overflow issue. Other time source are expensive here.
Should I limit the code only running on 64bit kernel?
Thanks,
-Aubrey
Quoted 7 lines
I misunderstood, you mean 32bit kernel, not 32bit machine. Theoretically 32bit kernel can use AVX512, but not sure if anyone use it like this.
get_jiffies_64()includesjiffies_lockops so not good in context switch. So I want to use rawjiffies_64here. jiffies is a good candidate but it has wraparound overflow issue. Other time source are expensive here.Should I limit the code only running on 64bit kernel?
Yes making it 64bit only should be fine.
Other alternative would be to use 32bit jiffies on 32bit. I assume wrapping is not that big a problem here.
-Andi
Quoted 8 lines · Andi Kleen
Quoted 7 lines
I misunderstood, you mean 32bit kernel, not 32bit machine. Theoretically 32bit kernel can use AVX512, but not sure if anyone use it like this.
get_jiffies_64()includesjiffies_lockops so not good in context switch. So I want to use rawjiffies_64here. jiffies is a good candidate but it has wraparound overflow issue. Other time source are expensive here.Should I limit the code only running on 64bit kernel?
Yes making it 64bit only should be fine.
I think I'd rather just disable AVX512 itself on 32-bit and be done with it. I think more than half of the ~2k of XSAVE space that it consumes in *every* *task* is just pure waste because it has to be 0's.
This ~2k of extra space is also lowmem, which makes it even more valuable.
Quoted 7 lines · Dave Hansen
Quoted 6 lines · Andi Kleen
Quoted 6 lines
I misunderstood, you mean 32bit kernel, not 32bit machine. Theoretically 32bit kernel can use AVX512, but not sure if anyone use it like this.
get_jiffies_64()includesjiffies_lockops so not good in context switch. So I want to use rawjiffies_64here. jiffies is a good candidate but it has wraparound overflow issue. Other time source are expensive here.
Quoted 8 lines
Quoted 2 lines
Quoted 1 line
Should I limit the code only running on 64bit kernel?
Yes making it 64bit only should be fine.
I think I'd rather just disable AVX512 itself on 32-bit and be done with it. I think more than half of the ~2k of XSAVE space that it consumes in *every* *task* is just pure waste because it has to be 0's.
This ~2k of extra space is also lowmem, which makes it even more valuable.
That will actually break programs.
If someone compiled binaries with -march=native on a system with AVX512 they wouldn't work anymore.
Don't think we can do it.
-Andi
Quoted 6 lines · Andi Kleen
Quoted 6 lines
I misunderstood, you mean 32bit kernel, not 32bit machine. Theoretically 32bit kernel can use AVX512, but not sure if anyone use it like this.
get_jiffies_64()includesjiffies_lockops so not good in context switch. So I want to use rawjiffies_64here. jiffies is a good candidate but it has wraparound overflow issue. Other time source are expensive here.
Quoted 9 lines
Quoted 1 line
Should I limit the code only running on 64bit kernel?
Yes making it 64bit only should be fine.
Other alternative would be to use 32bit jiffies on 32bit. I assume wrapping is not that big a problem here.
Thomas, is this acceptable?
Thanks,
-Aubrey
Quoted 7 lines · Aubrey
Quoted 6 lines · Andi Kleen
Quoted 6 lines
I misunderstood, you mean 32bit kernel, not 32bit machine. Theoretically 32bit kernel can use AVX512, but not sure if anyone use it like this.
get_jiffies_64()includesjiffies_lockops so not good in context switch. So I want to use rawjiffies_64here. jiffies is a good candidate but it has wraparound overflow issue. Other time source are expensive here.
Quoted 8 lines
Quoted 7 lines
Quoted 1 line
Should I limit the code only running on 64bit kernel?
Yes making it 64bit only should be fine.
Other alternative would be to use 32bit jiffies on 32bit. I assume wrapping is not that big a problem here.
Thomas, is this acceptable?
Just do the math. jiffies on 32bit wrap around depending on HZ:
HZ=100 248 days
HZ=1000 24 days
So, yes it takes quite some time, but from then on the information is bogus. Whether that matters or not is a different question. At least it needs proper documentation.
Thanks,
tglx
Quoted 3 lines · Thomas Gleixner
What exactly prevents a 32bit kernel from having the AVX512 feature bit set? And if it cannot be set on 32bit, then why are you compiling that code in at all?
There are three different AVX-512 states (and three bits) which Aubrey's patch checks. All three have different rules. Here's a summary along with some relevant SDM quotes from Vol1-13.6.
Opmask state: All opmask registers can be set in 32-bit mode. ZMM_Hi256 state: "An execution of XRSTOR or XRSTORS outside 64-bit mode
does not update ZMM8_H–ZMM15_H." This implies that ZMM0_H-ZMM7_H *are* updated in 32-bit mode.
Hi16_ZMM state: "Outside 64-bit mode, Hi16_ZMM state is always in its
initial configuration."
All of Hi16_ZMM and *part* of ZMM_Hi256 can not be practically used in 32-bit mode. But, even using part of ZMM_Hi256 means the xfeature bit will be set.
So, 2/3 of the features can be used in 32-bit mode. Nothing that I can find _prevents_ those features from being used in 32-bit mode.
Aubrey, do you have information to the contrary?
Quoted 16 lines · Dave Hansen
Quoted 3 lines · Thomas Gleixner
What exactly prevents a 32bit kernel from having the AVX512 feature bit set? And if it cannot be set on 32bit, then why are you compiling that code in at all?
There are three different AVX-512 states (and three bits) which Aubrey's patch checks. All three have different rules. Here's a summary along with some relevant SDM quotes from Vol1-13.6.
Opmask state: All opmask registers can be set in 32-bit mode.
ZMM_Hi256state: "An execution of XRSTOR or XRSTORS outside 64-bit modedoes not update ZMM8_H–ZMM15_H." This implies that ZMM0_H-ZMM7_H *are* updated in 32-bit mode.
Hi16_ZMMstate: "Outside 64-bit mode,Hi16_ZMMstate is always in its
initial configuration."
Quoted 9 lines
All of
Hi16_ZMMand *part* ofZMM_Hi256can not be practically used in 32-bit mode. But, even using part ofZMM_Hi256means the xfeature bit will be set.So, 2/3 of the features can be used in 32-bit mode. Nothing that I can find _prevents_ those features from being used in 32-bit mode.
Aubrey, do you have information to the contrary?
Thanks Dave for the summary, similar info I have from SDM, so yes, 32bit kernel can have xfeature bits set.