From f9ced08de30c37838756e8227bd091f80ad9cafa Mon Sep 17 00:00:00 2001 From: Ben Noordhuis Date: Thu, 24 Apr 2014 04:27:40 +0200 Subject: [PATCH] deps: make v8 use CLOCK_REALTIME_COARSE Date.now() indirectly calls gettimeofday() on Linux and that's a system call that is extremely expensive on virtualized systems when the host operating system has to emulate access to the hardware clock. Case in point: output from `perf record -c 10000 -e cycles:u -g -i` for a benchmark/http_simple bytes/8 benchmark with a light load of 50 concurrent clients: 53.69% node node [.] v8::internal::OS::TimeCurrentMillis() | --- v8::internal::OS::TimeCurrentMillis() | |--99.77%-- v8::internal::Runtime_DateCurrentTime(v8::internal::Arguments, v8::internal::Isolate*) | 0x23587880618e That's right - over half of user time spent inside the V8 function that calls gettimeofday(). Notably, nearly all system time gets attributed to acpi_pm_read(), the kernel function that reads the ACPI power management timer: 32.49% node [kernel.kallsyms] [k] acpi_pm_read | --- acpi_pm_read | |--98.40%-- __getnstimeofday | getnstimeofday | | | |--71.61%-- do_gettimeofday | | sys_gettimeofday | | system_call_fastpath | | 0x7fffbbaf6dbc | | | | | |--98.72%-- v8::internal::OS::TimeCurrentMillis() The cost of the gettimeofday() system call is normally measured in nanoseconds but we were seeing 100 us averages and spikes >= 1000 us. The numbers were so bad, my initial hunch was that the node process was continuously getting rescheduled inside the system call... v8::internal::OS::TimeCurrentMillis()'s most frequent caller is v8::internal::Runtime_DateCurrentTime(), the V8 run-time function that's behind Date.now(). The timeout handling logic in lib/http.js and lib/net.js calls into lib/timers.js and that module will happily call Date.now() hundreds or even thousands of times per second. If you saw exports._unrefActive() show up in --prof output a lot, now you know why. That's why this commit makes V8 switch over to clock_gettime() on Linux. In particular, it checks if CLOCK_REALTIME_COARSE is available and has a resolution <= 1 ms because in that case the clock_gettime() call can be fully serviced from the vDSO. It speeds up the aforementioned benchmark by about 100% on the affected systems and should go a long way toward addressing the latency issues that StrongLoop customers have been reporting. This patch will be upstreamed as a CR against V8 3.26. I'm sending it as a pull request for v0.10 first because that's what our users are running and because the delta between 3.26 and 3.14 is too big to reasonably back-port the patch. I'll open a pull request for the master branch once the CR lands upstream. Signed-off-by: Trevor Norris Signed-off-by: Fedor Indutny --- src/platform-posix.cc | 26 ++++++++++++++++++++++---- 1 file changed, 22 insertions(+), 4 deletions(-) diff --git a/src/platform-posix.cc b/src/platform-posix.cc index ad74eba..3c86868 100644 --- a/src/platform-posix.cc +++ b/src/platform-posix.cc @@ -188,19 +188,37 @@ int OS::GetUserTime(uint32_t* secs, uint32_t* usecs) { double OS::TimeCurrentMillis() { - struct timeval tv; - if (gettimeofday(&tv, NULL) < 0) return 0.0; - return (static_cast(tv.tv_sec) * 1000) + - (static_cast(tv.tv_usec) / 1000); + return static_cast(Ticks()) / 1000; } int64_t OS::Ticks() { +#if defined(__linux__) + static clockid_t clock_id = static_cast(-1); + struct timespec spec; + if (clock_id == static_cast(-1)) { + // CLOCK_REALTIME_COARSE may not be defined by the system headers but + // might still be supported by the kernel so use the clock id directly. + // Only use CLOCK_REALTIME_COARSE when its granularity <= 1 ms. + const clockid_t clock_realtime_coarse = 5; + if (clock_getres(clock_realtime_coarse, &spec) == 0 && + spec.tv_nsec <= 1000 * 1000) { + clock_id = clock_realtime_coarse; + } else { + clock_id = CLOCK_REALTIME; + } + } + if (clock_gettime(clock_id, &spec) != 0) { + return 0; // Not really possible. + } + return static_cast(spec.tv_sec) * 1000000 + (spec.tv_nsec / 1000); +#else // gettimeofday has microsecond resolution. struct timeval tv; if (gettimeofday(&tv, NULL) < 0) return 0; return (static_cast(tv.tv_sec) * 1000000) + tv.tv_usec; +#endif } -- 1.9.1