From 2aebc5a129c93498318a950fcb112be5de3e88f5 Mon Sep 17 00:00:00 2001 From: Vincent Fu Date: Mon, 19 Jun 2017 15:12:36 -0400 Subject: [PATCH] gettime: for better accuracy calculate cycles_per_msec instead of cycles_per_usec fio (before the nsec changes) would finish about 9 sec early for a 16 hour job when relying on the CPU clock. This is because fio calculates cycles_per_usec to carry out the clock ticks to time conversion. cycles_per_usec only provides 4 significant digits. Changing this to cycles_per_msec provides 7 significant digits and makes the actual job run time more closely match wall time. --- gettime.c | 22 +++++++++++----------- 1 file changed, 11 insertions(+), 11 deletions(-) diff --git a/gettime.c b/gettime.c index 763e5744..093b6a49 100644 --- a/gettime.c +++ b/gettime.c @@ -15,7 +15,7 @@ #if defined(ARCH_HAVE_CPU_CLOCK) #ifndef ARCH_CPU_CLOCK_CYCLES_PER_USEC -static unsigned long cycles_per_usec; +static unsigned long cycles_per_msec; static unsigned long long cycles_start; static unsigned long long clock_mult; static unsigned long long max_cycles_mask; @@ -231,7 +231,7 @@ void fio_gettime(struct timespec *tp, void fio_unused *caller) } #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC) -static unsigned long get_cycles_per_usec(void) +static unsigned long get_cycles_per_msec(void) { struct timespec s, e; uint64_t c_s, c_e; @@ -257,7 +257,7 @@ static unsigned long get_cycles_per_usec(void) } while (1); fio_clock_source = old_cs; - return (c_e - c_s) / elapsed; + return (c_e - c_s) * 1000 / elapsed; } #define NR_TIME_ITERS 50 @@ -269,10 +269,10 @@ static int calibrate_cpu_clock(void) int i, samples, sft = 0; unsigned long long tmp, max_ticks, max_mult; - cycles[0] = get_cycles_per_usec(); + cycles[0] = get_cycles_per_msec(); S = delta = mean = 0.0; for (i = 0; i < NR_TIME_ITERS; i++) { - cycles[i] = get_cycles_per_usec(); + cycles[i] = get_cycles_per_msec(); delta = cycles[i] - mean; if (delta) { mean += delta / (i + 1.0); @@ -309,13 +309,13 @@ static int calibrate_cpu_clock(void) dprint(FD_TIME, "cycles[%d]=%llu\n", i, (unsigned long long) cycles[i]); avg /= samples; - cycles_per_usec = avg; + cycles_per_msec = avg; dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg); dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n", (unsigned long long) minc, (unsigned long long) maxc, mean, S); - max_ticks = MAX_CLOCK_SEC * cycles_per_usec * 1000000ULL; + max_ticks = MAX_CLOCK_SEC * cycles_per_msec * 1000ULL; max_mult = ULLONG_MAX / max_ticks; dprint(FD_TIME, "\n\nmax_ticks=%llu, __builtin_clzll=%d, max_mult=%llu\n", max_ticks, __builtin_clzll(max_ticks), max_mult); @@ -324,7 +324,7 @@ static int calibrate_cpu_clock(void) * Find the largest shift count that will produce * a multiplier that does not exceed max_mult */ - tmp = max_mult * cycles_per_usec / 1000; + tmp = max_mult * cycles_per_msec / 1000000; while (tmp > 1) { tmp >>= 1; sft++; @@ -332,19 +332,19 @@ static int calibrate_cpu_clock(void) } clock_shift = sft; - clock_mult = (1ULL << sft) * 1000 / cycles_per_usec; + clock_mult = (1ULL << sft) * 1000000 / cycles_per_msec; dprint(FD_TIME, "clock_shift=%u, clock_mult=%llu\n", clock_shift, clock_mult); // Find the greatest power of 2 clock ticks that is less than the ticks in MAX_CLOCK_SEC_2STAGE max_cycles_shift = max_cycles_mask = 0; - tmp = MAX_CLOCK_SEC * 1000000ULL * cycles_per_usec; + tmp = MAX_CLOCK_SEC * 1000ULL * cycles_per_msec; dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, max_cycles_shift); while (tmp > 1) { tmp >>= 1; max_cycles_shift++; dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, max_cycles_shift); } - // if use use (1ULL << max_cycles_shift) * 1000 / cycles_per_usec here we will + // if use use (1ULL << max_cycles_shift) * 1000 / cycles_per_msec here we will // have a discontinuity every (1ULL << max_cycles_shift) cycles nsecs_for_max_cycles = ((1ULL << max_cycles_shift) * clock_mult) >> clock_shift; -- 2.25.1