CellBench provides the ability to measure the running time of pipelines. This is done using the time_methods()
function which runs in the same way that apply_methods()
does, with the difference that it does not run in parallel. This is an intentional design choice because running things in parallel usually results in some competition for computer resource and therefore produces less reliable or stable timings.
The setup for timing methods is identical to applying methods. You have a list of data and a list of functions, then you use time_methods()
instead of apply_methods()
.
library(CellBench)
# wrap a simple vector in a list
datasets <- list(
data1 = c(1, 2, 3)
)
# use Sys.sleep in functions to simulate long-running functions
transform <- list(
log = function(x) { Sys.sleep(0.1); log(x) },
sqrt = function(x) { Sys.sleep(0.1); sqrt(x) }
)
# time the functions
res <- datasets %>%
time_methods(transform)
res
## # A tibble: 2 × 3
## data transform timed_result
## <fct> <fct> <list>
## 1 data1 log <named list [2]>
## 2 data1 sqrt <named list [2]>
Where we usually have the result
column we now have timed_result
, this is a list of two objects: the timing object and the result. It is necessary to keep the result so that we can chain computations together.
res$timed_result[[1]]
## $timing
## user system elapsed
## 0.001 0.000 0.101
##
## $result
## [1] 0.0000000 0.6931472 1.0986123
As is the case with apply_methods()
, more lists of methods can be applied and results will expand out combinatorially. The timings in this case will be cumulative over the methods applied.
transform2 <- list(
plus = function(x) { Sys.sleep(0.1); x + 1 },
minus = function(x) { Sys.sleep(0.1); x - 1 }
)
res2 <- datasets %>%
time_methods(transform) %>%
time_methods(transform2)
res2
## # A tibble: 4 × 4
## data transform transform2 timed_result
## <fct> <fct> <fct> <list>
## 1 data1 log plus <named list [2]>
## 2 data1 log minus <named list [2]>
## 3 data1 sqrt plus <named list [2]>
## 4 data1 sqrt minus <named list [2]>
The class of results from time_methods()
is benchmark_timing_tbl
. Once all methods have been applied, the result may be discarded using unpack_timing()
and the object can be transformed into a more flat tbl
representation. See ?proc_time
for an explanation of what user
, system
and elapsed
refer to.
The timing values have been converted to Duration
objects from the lubridate
package, these behave as numeric measurements in seconds but have nicer printing properties (try lubridate::duration(1000, units = "seconds")
).
# discard results and expand out timings into columns
res2 %>%
unpack_timing()
## # A tibble: 4 × 6
## data transform transform2 user system elapsed
## <fct> <fct> <fct> <Duration> <Duration> <Duration>
## 1 data1 log plus 0.003s 0s 0.203s
## 2 data1 log minus 0.003s 0s 0.203s
## 3 data1 sqrt plus 0.006s 0s 0.205s
## 4 data1 sqrt minus 0.006s 0s 0.205s
Alternatively the timing information can be discarded and a benchmark_tbl
can be produced using strip_timing()
.
# discard timings and produce benchmark_tbl object
res2 %>%
strip_timing()
## # A tibble: 4 × 4
## data transform transform2 result
## <fct> <fct> <fct> <list>
## 1 data1 log plus <dbl [3]>
## 2 data1 log minus <dbl [3]>
## 3 data1 sqrt plus <dbl [3]>
## 4 data1 sqrt minus <dbl [3]>
CellBench provides a simple way to measure the running times of pipelines from various combinations of methods. This is done with the time_methods()
function which is called in the same way as apply_methods()
and has the same chaining properties. The resultant object can be transformed in two useful ways, as a flat tibble
with timings expanded out as columns and discarding the results, or as a benchmark_tbl
with the results as a list-column
and discarding the timings.