On 3/28/2024 1:06 AM, Mingwei Zhang wrote:
On Wed, Mar 27, 2024, Mi, Dapeng wrote:On 3/27/2024 1:36 PM, Mingwei Zhang wrote:On Wed, Jan 03, 2024, Dapeng Mi wrote:When running pmu test on SPR, sometimes the following failure is reported. PMU version: 2 GP counters: 8 GP counter width: 48 Mask length: 8 Fixed counters: 3 Fixed counter width: 48 1000000 <= 55109398 <= 50000000 FAIL: Intel: core cycles-0 1000000 <= 18279571 <= 50000000 PASS: Intel: core cycles-1 1000000 <= 12238092 <= 50000000 PASS: Intel: core cycles-2 1000000 <= 7981727 <= 50000000 PASS: Intel: core cycles-3 1000000 <= 6984711 <= 50000000 PASS: Intel: core cycles-4 1000000 <= 6773673 <= 50000000 PASS: Intel: core cycles-5 1000000 <= 6697842 <= 50000000 PASS: Intel: core cycles-6 1000000 <= 6747947 <= 50000000 PASS: Intel: core cycles-7 The count of the "core cycles" on first counter would exceed the upper boundary and leads to a failure, and then the "core cycles" count would drop gradually and reach a stable state. That looks reasonable. The "core cycles" event is defined as the 1st event in xxx_gp_events[] array and it is always verified at first. when the program loop() is executed at the first time it needs to warm up the pipeline and cache, such as it has to wait for cache is filled. All these warm-up work leads to a quite large core cycles count which may exceeds the verification range. The event "instructions" instead of "core cycles" is a good choice as the warm-up event since it would always return a fixed count. Thus switch instructions and core cycles events sequence in the xxx_gp_events[] array.The observation is great. However, it is hard to agree that we fix the problem by switching the order. Maybe directly tweaking the N from 50 to a larger value makes more sense. Thanks. -Mingweiyeah, a larger upper boundary can fix the fault as well, but the question is how large it would be enough. For different CPU model, the needed cycles could be different for warming up. So we may have to set a quite large upper boundary but a large boundary would decrease credibility of this validation. Not sure which one is better. Any inputs from other ones?Just checked with an expert from our side, so "core cycles" (0x003c) is affected the current CPU state/frequency, ie., its counting value could vary largely. In that sense, "warming" up seems reasonable. However, switching the order would be a terrible idea for maintanence since people will forget it and the problem will come back. From another perspective, "warming" up seems just a best effort. Nobody knows how warm is really warm. Besides, some systems might turn off some C-State and may set a cap on max turbo frequency. All of these will directly affect the warm-up process and the counting result of 0x003c. So, while adding a warm-up blob is reasonable, tweaking the heuristics seems to be same for me. Regarding the value, I think I will rely on your experiments and observation.
Per my understanding, most of extra cpu cycles should come from the warm up for cache. If we don't want to change the validation order, it may be doable to add an extra warm-up phase before starting the validation. Thus we don't need to enlarge the upper boundary. It looks not a preferred way since it would decrease the credibility of the validation.
Let me try to add a warm-up phase first and check if it works as expect.
Thanks. -MingweiSigned-off-by: Dapeng Mi <dapeng1.mi@xxxxxxxxxxxxxxx> --- x86/pmu.c | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/x86/pmu.c b/x86/pmu.c index a42fff8d8b36..67ebfbe55b49 100644 --- a/x86/pmu.c +++ b/x86/pmu.c @@ -31,16 +31,16 @@ struct pmu_event { int min; int max; } intel_gp_events[] = { - {"core cycles", 0x003c, 1*N, 50*N}, {"instructions", 0x00c0, 10*N, 10.2*N}, + {"core cycles", 0x003c, 1*N, 50*N}, {"ref cycles", 0x013c, 1*N, 30*N}, {"llc references", 0x4f2e, 1, 2*N}, {"llc misses", 0x412e, 1, 1*N}, {"branches", 0x00c4, 1*N, 1.1*N}, {"branch misses", 0x00c5, 0, 0.1*N}, }, amd_gp_events[] = { - {"core cycles", 0x0076, 1*N, 50*N}, {"instructions", 0x00c0, 10*N, 10.2*N}, + {"core cycles", 0x0076, 1*N, 50*N}, {"branches", 0x00c2, 1*N, 1.1*N}, {"branch misses", 0x00c3, 0, 0.1*N}, }, fixed_events[] = { @@ -307,7 +307,7 @@ static void check_counter_overflow(void) int i; pmu_counter_t cnt = { .ctr = MSR_GP_COUNTERx(0), - .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[1].unit_sel /* instructions */, + .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[0].unit_sel /* instructions */, }; overflow_preset = measure_for_overflow(&cnt); @@ -365,11 +365,11 @@ static void check_gp_counter_cmask(void) { pmu_counter_t cnt = { .ctr = MSR_GP_COUNTERx(0), - .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[1].unit_sel /* instructions */, + .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[0].unit_sel /* instructions */, }; cnt.config |= (0x2 << EVNTSEL_CMASK_SHIFT); measure_one(&cnt); - report(cnt.count < gp_events[1].min, "cmask"); + report(cnt.count < gp_events[0].min, "cmask"); } static void do_rdpmc_fast(void *ptr) @@ -446,7 +446,7 @@ static void check_running_counter_wrmsr(void) uint64_t count; pmu_counter_t evt = { .ctr = MSR_GP_COUNTERx(0), - .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[1].unit_sel, + .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[0].unit_sel, }; report_prefix_push("running counter wrmsr"); @@ -455,7 +455,7 @@ static void check_running_counter_wrmsr(void) loop(); wrmsr(MSR_GP_COUNTERx(0), 0); stop_event(&evt); - report(evt.count < gp_events[1].min, "cntr"); + report(evt.count < gp_events[0].min, "cntr"); /* clear status before overflow test */ if (this_cpu_has_perf_global_status()) @@ -493,7 +493,7 @@ static void check_emulated_instr(void) pmu_counter_t instr_cnt = { .ctr = MSR_GP_COUNTERx(1), /* instructions */ - .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[1].unit_sel, + .config = EVNTSEL_OS | EVNTSEL_USR | gp_events[0].unit_sel, }; report_prefix_push("emulated instruction"); -- 2.34.1
