test_cpu_time.cpp

Go to the documentation of this file.

//    OpenVPN -- An application to securely tunnel IP networks
//               over a single port, with support for SSL/TLS-based
//               session authentication and key exchange,
//               packet encryption, packet authentication, and
//               packet compression.
//
//    Copyright (C) 2019- OpenVPN Inc.
//
//    SPDX-License-Identifier: MPL-2.0 OR AGPL-3.0-only WITH openvpn3-openssl-exception
//
 
 
// We have two set of measurements for these tests
//
// 1. A coarse measurement based on time()
//    These are tracked in 'chk_start' and 'chk_end'
//
// 2. A fine grained measurement from cpu_time()
//    These are tracked in 'start' and 'end'
//
// We calculate the difference before and after a
// a workload has run, to measure how long it ran.
// This is done for both measurement approaches.
// The runtime is saved in runtime and chk_runtime
//
// To pass this test, absolute difference between
// runtime and chk_runtime must be less than 1 second.
//
 
// #define DEBUG  // Define this macro to get more details
 
#include "test_common.hpp"
#include <cstdint>
#include <unistd.h>
#include <memory>
#include <mutex>
#include <openvpn/time/cputime.hpp>
#include <algorithm>
 
#ifdef DEBUG
#define DEBUG_DUMP(msg, st, en, rt, chst, chen, chrt, md)              \
    std::cout << msg << std::endl                                      \
              << "start = " << std::to_string(st) << std::endl         \
              << "end = " << std::to_string(en) << std::endl           \
              << "runtime = " << std::to_string(rt) << std::endl       \
              << "chk_start = " << std::to_string(chst) << std::endl   \
              << "chk_end = " << std::to_string(chen) << std::endl     \
              << "chk_runtime = " << std::to_string(chrt) << std::endl \
              << "measurement difference = "                           \
              << std::to_string(md) << std::endl                       \
              << "--------------------------------------" << std::endl
 
#else
#define DEBUG_DUMP(msg, st, en, rt, chst, chen, chrt, md) \
    {                                                     \
    }
#endif
 
#define MEASURE(v, chkv, thread)          \
    double v = openvpn::cpu_time(thread); \
    ASSERT_GE(v, 0);                      \
    double chkv = static_cast<double>(time(NULL))
 
#define CALCULATE(msg, st, en, rt, chst, chen, chrt, md) \
    double rt = en - st;                                 \
    double chrt = chen - chst;                           \
    double md = std::max(rt, chrt) - std::min(rt, chrt); \
    DEBUG_DUMP(msg, st, en, rt, chst, chen, chrt, md)
 
 
typedef std::shared_ptr<std::thread> ThreadPtr;
 
 
// For simplicty, keep the total thread runtime
// in a global variable, protected by a mutex
// on updates
std::mutex update_guard;
double thread_runtime = 0;
 
void update_thread_runtime(double val)
{
    std::lock_guard<std::mutex> ug(update_guard);
    thread_runtime += val;
}
 
 
namespace unittests {
static void workload(const uint16_t multiplier)
{
    // A very simple busy loop workload
    //
    // We can't use sleep() or any similar timing
    // as this does not increase the tracked runtime
    // in the kernel; the process does not really run.
    //
 
    std::random_device rd;
    std::mt19937 gen(rd()); // Standard mersenne_twister_engine seeded with rd()
 
 
    double d = 0;
    for (unsigned int i = UINT16_MAX * multiplier; i > 0; i--)
    {
        d += static_cast<double>(gen());
    }
    (void)d;
}
 
 
TEST(CPUTime, cpu_time_pid)
{
    // Measure the runtime of the workload
    MEASURE(start, chk_start, false);
    workload(400);
    MEASURE(end, chk_end, false);
 
    // Calculate runtimes and differences
    CALCULATE("single PID",
              start,
              end,
              runtime,
              chk_start,
              chk_end,
              chk_runtime,
              measurement_diff);
 
    ASSERT_LT(measurement_diff, 10);
}
 
 
void worker_thread(const uint8_t id)
{
    MEASURE(thr_start, chk_thr_start, true);
    workload(400);
    MEASURE(thr_end, chk_thr_end, true);
 
    CALCULATE("Worker thread " << std::to_string(id),
              thr_start,
              thr_end,
              thr_runtime,
              chk_thr_start,
              chk_thr_end,
              chk_thr_runtime,
              thr_measurement_diff);
    update_thread_runtime(thr_runtime);
 
    // Since the chk_thr_runtime (chk_thr_end - chk_thr_start)
    // is based on epoc seconds of the system, this doesn't
    // give a too good number when running multiple threads.
    //
    // If more threads are running on the same CPU core,
    // one of the threads might be preempted.  The clock time
    // (chk_thr_runtime) will continue to tick, but the real
    // runtime (thr_runtime) will not.  Which will increase
    // the difference between the measured runtimes.
    //
    // The value of 5 is just an educated guess of what
    // we might find acceptable.  This might be too high
    // on an idle system, but too low on a loaded system.
    //
    ASSERT_LT(thr_measurement_diff, 5);
}
 
 
void run_threads(const uint8_t num_threads)
{
    std::vector<ThreadPtr> threads;
    for (uint8_t i = 0; i < num_threads; i++)
    {
        ThreadPtr tp;
        tp = std::make_shared<std::thread>([id = i]()
                                           { worker_thread(id); });
        threads.push_back(std::move(tp));
    }
 
    for (const auto &t : threads)
    {
        t->join();
    }
}
 
 
TEST(CPUTime, cpu_time_thread_1)
{
    // Meassure running a single worker thread
    MEASURE(parent_start, chk_parent_start, false);
    run_threads(1);
    MEASURE(parent_end, chk_parent_end, false);
 
    CALCULATE("Parent thread - 1 child thread",
              parent_start,
              parent_end,
              runtime,
              chk_parent_start,
              chk_parent_end,
              chk_runtime,
              parent_diff);
 
    ASSERT_LT(parent_diff, 10);
}
 
 
TEST(CPUTime, cpu_time_thread_numcores)
{
    // Use number of available cores
    auto num_cores = std::min(std::thread::hardware_concurrency(), 1u);
 
    // Meassure running a single worker thread
    MEASURE(parent_start, chk_parent_start, false);
    run_threads(static_cast<uint8_t>(num_cores));
    MEASURE(parent_end, chk_parent_end, false);
 
    CALCULATE("Parent thread - " << std::to_string(num_cores) << " child thread",
              parent_start,
              parent_end,
              runtime,
              chk_parent_start,
              chk_parent_end,
              chk_runtime,
              parent_diff);
#ifdef DEBUG
    std::cout << "Total thread runtime: " << std::to_string(thread_runtime) << std::endl;
#endif
 
    // The main process (this PID) will have a total runtime
    // which accounts for all runtime of the running threads.
    // We still give a bit extra slack, to reduce the risk of
    // false positives, due to the possibility of premption
    // (see comment in worker_thread() for details).  But
    // the difference should not neccesarily deviate as much
    // here.
    ASSERT_LT(parent_diff, 3 + thread_runtime);
}
} // namespace unittests