ceph/src/common/Cycles.cc

   1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
   2 // vim: ts=8 sw=2 smarttab
   3 /*
   4  * Ceph - scalable distributed file system
   5  *
   6  * Copyright (C) 2014 UnitedStack <haomai@unitedstack.com>
   7  *
   8  * Author: Haomai Wang <haomaiwang@gmail.com>
   9  *
  10  * This is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License version 2.1, as published by the Free Software
  13  * Foundation.  See file COPYING.
  14  *
  15  */
  16 /* Copyright (c) 2011-2014 Stanford University
  17  *
  18  * Permission to use, copy, modify, and distribute this software for any
  19  * purpose with or without fee is hereby granted, provided that the above
  20  * copyright notice and this permission notice appear in all copies.
  21  *
  22  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR(S) DISCLAIM ALL WARRANTIES
  23  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
  24  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL AUTHORS BE LIABLE FOR
  25  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
  26  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
  27  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
  28  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
  29  */
  30
  31 #include "debug.h"
  32 #include "Cycles.h"
  33
  34 double Cycles::cycles_per_sec = 0;
  35
  36 /**
  37  * Perform once-only overall initialization for the Cycles class, such
  38  * as calibrating the clock frequency.  This method must be called
  39  * before using the Cycles module.
  40  *
  41  * It is not initialized by default because the timing loops cause
  42  * general process startup times to balloon
  43  * (http://tracker.ceph.com/issues/15225).
  44  */
  45 void Cycles::init()
  46 {
  47   if (cycles_per_sec != 0)
  48     return;
  49
  50   // Skip initialization if rtdsc is not implemented
  51   if (rdtsc() == 0)
  52     return;
  53
  54   // Compute the frequency of the fine-grained CPU timer: to do this,
  55   // take parallel time readings using both rdtsc and gettimeofday.
  56   // After 10ms have elapsed, take the ratio between these readings.
  57
  58   struct timeval start_time, stop_time;
  59   uint64_t micros;
  60   double old_cycles;
  61
  62   // There is one tricky aspect, which is that we could get interrupted
  63   // between calling gettimeofday and reading the cycle counter, in which
  64   // case we won't have corresponding readings.  To handle this (unlikely)
  65   // case, compute the overall result repeatedly, and wait until we get
  66   // two successive calculations that are within 0.1% of each other.
  67   old_cycles = 0;
  68   while (1) {
  69     if (gettimeofday(&start_time, NULL) != 0) {
  70       ceph_abort_msg("couldn't read clock");
  71     }
  72     uint64_t start_cycles = rdtsc();
  73     while (1) {
  74       if (gettimeofday(&stop_time, NULL) != 0) {
  75         ceph_abort_msg("couldn't read clock");
  76       }
  77       uint64_t stop_cycles = rdtsc();
  78       micros = (stop_time.tv_usec - start_time.tv_usec) +
  79           (stop_time.tv_sec - start_time.tv_sec)*1000000;
  80       if (micros > 10000) {
  81         cycles_per_sec = static_cast<double>(stop_cycles - start_cycles);
  82         cycles_per_sec = 1000000.0*cycles_per_sec/ static_cast<double>(micros);
  83         break;
  84       }
  85     }
  86     double delta = cycles_per_sec/1000.0;
  87     if ((old_cycles > (cycles_per_sec - delta)) &&
  88         (old_cycles < (cycles_per_sec + delta))) {
  89       return;
  90     }
  91     old_cycles = cycles_per_sec;
  92   }
  93 }
  94
  95 /**
  96  * Return the number of CPU cycles per second.
  97  */
  98 double Cycles::per_second()
  99 {
 100   return get_cycles_per_sec();
 101 }
 102
 103 /**
 104  * Given an elapsed time measured in cycles, return a floating-point number
 105  * giving the corresponding time in seconds.
 106  * \param cycles
 107  *      Difference between the results of two calls to rdtsc.
 108  * \param cycles_per_sec
 109  *      Optional parameter to specify the frequency of the counter that #cycles
 110  *      was taken from. Useful when converting a remote machine's tick counter
 111  *      to seconds. The default value of 0 will use the local processor's
 112  *      computed counter frequency.
 113  * \return
 114  *      The time in seconds corresponding to cycles.
 115  */
 116 double Cycles::to_seconds(uint64_t cycles, double cycles_per_sec)
 117 {
 118   if (cycles_per_sec == 0)
 119     cycles_per_sec = get_cycles_per_sec();
 120   return static_cast<double>(cycles)/cycles_per_sec;
 121 }
 122
 123 /**
 124  * Given a time in seconds, return the number of cycles that it
 125  * corresponds to.
 126  * \param seconds
 127  *      Time in seconds.
 128  * \param cycles_per_sec
 129  *      Optional parameter to specify the frequency of the counter that #cycles
 130  *      was taken from. Useful when converting a remote machine's tick counter
 131  *      to seconds. The default value of 0 will use the local processor's
 132  *      computed counter frequency.
 133  * \return
 134  *      The approximate number of cycles corresponding to #seconds.
 135  */
 136 uint64_t Cycles::from_seconds(double seconds, double cycles_per_sec)
 137 {
 138   if (cycles_per_sec == 0)
 139     cycles_per_sec = get_cycles_per_sec();
 140   return (uint64_t) (seconds*cycles_per_sec + 0.5);
 141 }
 142
 143 /**
 144  * Given an elapsed time measured in cycles, return an integer
 145  * giving the corresponding time in microseconds. Note: to_seconds()
 146  * is faster than this method.
 147  * \param cycles
 148  *      Difference between the results of two calls to rdtsc.
 149  * \param cycles_per_sec
 150  *      Optional parameter to specify the frequency of the counter that #cycles
 151  *      was taken from. Useful when converting a remote machine's tick counter
 152  *      to seconds. The default value of 0 will use the local processor's
 153  *      computed counter frequency.
 154  * \return
 155  *      The time in microseconds corresponding to cycles (rounded).
 156  */
 157 uint64_t Cycles::to_microseconds(uint64_t cycles, double cycles_per_sec)
 158 {
 159   return to_nanoseconds(cycles, cycles_per_sec) / 1000;
 160 }
 161
 162 /**
 163  * Given an elapsed time measured in cycles, return an integer
 164  * giving the corresponding time in nanoseconds. Note: to_seconds()
 165  * is faster than this method.
 166  * \param cycles
 167  *      Difference between the results of two calls to rdtsc.
 168  * \param cycles_per_sec
 169  *      Optional parameter to specify the frequency of the counter that #cycles
 170  *      was taken from. Useful when converting a remote machine's tick counter
 171  *      to seconds. The default value of 0 will use the local processor's
 172  *      computed counter frequency.
 173  * \return
 174  *      The time in nanoseconds corresponding to cycles (rounded).
 175  */
 176 uint64_t Cycles::to_nanoseconds(uint64_t cycles, double cycles_per_sec)
 177 {
 178   if (cycles_per_sec == 0)
 179     cycles_per_sec = get_cycles_per_sec();
 180   return (uint64_t) (1e09*static_cast<double>(cycles)/cycles_per_sec + 0.5);
 181 }
 182
 183 /**
 184  * Given a number of nanoseconds, return an approximate number of
 185  * cycles for an equivalent time length.
 186  * \param ns
 187  *      Number of nanoseconds.
 188  * \param cycles_per_sec
 189  *      Optional parameter to specify the frequency of the counter that #cycles
 190  *      was taken from. Useful when converting a remote machine's tick counter
 191  *      to seconds. The default value of 0 will use the local processor's
 192  *      computed counter frequency.
 193  * \return
 194  *      The approximate number of cycles for the same time length.
 195  */
 196 uint64_t
 197 Cycles::from_nanoseconds(uint64_t ns, double cycles_per_sec)
 198 {
 199   if (cycles_per_sec == 0)
 200     cycles_per_sec = get_cycles_per_sec();
 201   return (uint64_t) (static_cast<double>(ns)*cycles_per_sec/1e09 + 0.5);
 202 }
 203
 204 /**
 205  * Busy wait for a given number of microseconds.
 206  * Callers should use this method in most reasonable cases as opposed to
 207  * usleep for accurate measurements. Calling usleep may put the the processor
 208  * in a low power mode/sleep state which reduces the clock frequency.
 209  * So, each time the process/thread wakes up from usleep, it takes some time
 210  * to ramp up to maximum frequency. Thus meausrements often incur higher
 211  * latencies.
 212  * \param us
 213  *      Number of microseconds.
 214  */
 215 void
 216 Cycles::sleep(uint64_t us)
 217 {
 218   uint64_t stop = Cycles::rdtsc() + Cycles::from_nanoseconds(1000*us);
 219   while (Cycles::rdtsc() < stop);
 220 }