Module Time_stamp_counter
High-performance timing.
This module provides the fast function now ()
which is our best effort high-performance cycle counter for a given platform. For x86 systems this retrieves the CPU's internal time stamp counter using the RDTSC instruction. For systems that do not have a RDTSC instruction, we fallback to using clock_gettime(CLOCK_MONOTONIC)
.
Here is a benchmark of execution time in nanos and allocations in words:
Name Time/Run mWd/Run ---------------------------- ---------- --------- Time.now 27.99ns 2.00w Time_ns.now 25.21ns TSC.Calibrator.calibrate 68.61ns TSC.now 6.87ns TSC.to_time 4.30ns 2.00w TSC.to_time (TSC.now ()) 8.75ns 2.00w TSC.to_time_ns 4.70ns TSC.to_time_ns(TSC.now ()) 9.56ns id 2.86ns TSC.Span.of_ns 11.66ns TSC.Span.to_ns 3.84ns
Type t
is an Int63.t
and consequently has no allocation overhead (on 64-bit machines), unlike Time.now ()
which returns a boxed float.
Functions are also provided to estimate the relationship of CPU time-stamp-counter frequency to real time, thereby allowing one to convert from t
to Time.t
. There are some caveats to this that are worth noting:
- The conversion to
Time.t
depends on an estimate of the time-stamp-counter frequency. This frequency may be volatile on some systems, thereby reducing the utility of this conversion. See theCalibrator
module below for details.
- The captured
t
can only be converted to aTime.t
if one also has a recently calibratedCalibrator.t
from the same machine.
- Put another way, it would not make sense to send a sexp of
t
from one box to another and then convert it to aTime.t
, becauset
counts the number of cycles since reset. So the measure only makes sense in the context of a single machine.
- Note that a cursory search for information about time stamp counter usage may give a false impression of its unreliability. Early processor implementations of TSC could be skewed by clock frequency changes (C-states) and by small differences between the startup time of each processor on a multi-processor machine. Modern hardware can usually be assumed to have an "invariant" tsc, and Linux has support to synchronize the initial counters at boot time when multiple processors are present.
type t
= private Core_kernel.Int63.t
include Bin_prot.Binable.S with type t := t
include Bin_prot.Binable.S_only_functions with type t := t
val bin_size_t : t Bin_prot.Size.sizer
val bin_write_t : t Bin_prot.Write.writer
val bin_read_t : t Bin_prot.Read.reader
val __bin_read_t__ : (int -> t) Bin_prot.Read.reader
This function only needs implementation if
t
exposed to be a polymorphic variant. Despite what the type reads, this does *not* produce a function after reading; instead it takes the constructor tag (int) before reading and reads the rest of the variantt
afterwards.
val bin_shape_t : Bin_prot.Shape.t
val bin_writer_t : t Bin_prot.Type_class.writer
val bin_reader_t : t Bin_prot.Type_class.reader
val bin_t : t Bin_prot.Type_class.t
include Ppx_sexp_conv_lib.Sexpable.S with type t := t
val t_of_sexp : Sexplib0.Sexp.t -> t
val sexp_of_t : t -> Sexplib0.Sexp.t
module Calibrator : sig ... end with type tsc := t
A calibrator contains a snapshot of machine-specific information that is used to convert between TSC values and clock time. This information needs to be calibrated periodically such that it stays updated w.r.t. changes in the CPU's time-stamp-counter frequency, which can vary depending on load, heat, etc. (Also see the comment in the
.ml
file.)
module Span : sig ... end
Span
indicates some integer number of cycles.
val now : unit -> t
val diff : t -> t -> Span.t
val add : t -> Span.t -> t
val to_int63 : t -> Core_kernel.Int63.t
val zero : t
val calibrator : Calibrator.t Core_kernel.Lazy.t
A default calibrator for the current process. Most programs can just use this calibrator; use others if collecting data from other processes / machines.
The first time this lazy value is forced, it spends approximately 3ms calibrating.
While the
Async
scheduler is running, this value is recalibrated regularly.
val to_time : t -> calibrator:Calibrator.t -> Core_kernel.Time.t
It is guaranteed that repeated calls will return nondecreasing
Time.t
values.
val to_time_ns : t -> calibrator:Calibrator.t -> Core_kernel.Time_ns.t