Bechamel.Analyze
Analyze module.
Micro-benchmark usually uses a linear-regression to estimates the execution time of a code segments. For example, the following table might represent {!Measurement_raw.t} array
collected by Benchmark.run
:
+-----+------+ | run | time | +-----+------+ | 1 | 19 | | 2 | 25 | | 3 | 37 | | 4 | 47 | | 5 | 56 | +-----+------+
Bechamel records 3000 samples and the number of iterations can grows geometrically (see Benchmark.run
). Then, Bechamel can use 2 algorithms:
The user can choose one of it. Currently, OLS
is the best to use. These algorithms will estimate the actual execution time of the code segment. Using OLS
with the above data would yield an estimated execution time of 9.6
nanoseconds with a goodness of fit (r²
) of 0.992
.
More generally, Bechamel lets the user choose the predictors and responder. Indeed, the user can use others metrics (such as perf
) and the API allows to analyze such metrics together.
module OLS : sig ... end
module RANSAC : sig ... end
ols ~r_square ~bootstrap ~predictors
is an Ordinary Least Square analysis on predictors
. It calculates r²
if r_square = true
. bootstrap
defines how many times Bechamel tries to resample measurements.
val one : 'a t -> Measure.witness -> Benchmark.t -> 'a
one analysis measure { Benchmark.stat; lr; kde; }
estimates the actual given measure
for one predictor
. So, one analysis time { Benchmark.stat; lr; kde; }
wants to estimate actual run-time
(or execution time) value, where analysis
is initialized with run
predictor.
val all :
'a t ->
Measure.witness ->
(string, Benchmark.t) Stdlib.Hashtbl.t ->
(string, 'a) Stdlib.Hashtbl.t
all analysis measure tbl
is an application of one
for all results from the given tbl
.
val merge :
'a t ->
Measure.witness list ->
(string, 'a) Stdlib.Hashtbl.t list ->
(string, (string, 'a) Stdlib.Hashtbl.t) Stdlib.Hashtbl.t
merge witnesses tbls
returns a dictionary where the key is the label of a measure (from the given witnesses
) and the value is the result of this specific measure.