tests: try to solve false negative on execution time benchmark

- new tunning of the limit threshold (Fail on 4ms RunModel with a gap of 0.25
- move benchmark test from check to a specific job and only for patched version

Refs #136

.gitlab-ci.yml

@@ -28,13 +28,6 @@ default:
   script:
   - R -e 'rcmdcheck::rcmdcheck(args = ifelse(as.logical(Sys.getenv("NOT_CRAN")), "", "--as-cran"), error_on = "warning")'
   - R -e 'install.packages(c("microbenchmark", "airGR"), repos = "http://cran.r-project.org")'
-  - Rscript tests/scheduled_tests/benchmarkRunModel.R
-  - R CMD INSTALL .
-  - Rscript tests/scheduled_tests/benchmarkRunModel.R
-  artifacts:
-    paths:
-      - tests/tmp/benchmark.tsv
-      - tests/tmp/mean_execution_time.tsv
 
 .check_not_cran:
   variables:
@@ -46,6 +39,19 @@ default:
     NOT_CRAN: "false"
   extends: .check
 
+benchmark_patched:
+  stage: check
+  variables:
+    R_VERSION: "patched"
+  script:
+  - Rscript tests/scheduled_tests/benchmarkRunModel.R
+  - R CMD INSTALL .
+  - Rscript tests/scheduled_tests/benchmarkRunModel.R
+  artifacts:
+    paths:
+      - tests/tmp/benchmark.tsv
+      - tests/tmp/mean_execution_time.tsv
+
 scheduled_tests_patched:
   # only:
   #   refs:

tests/testthat/benchmarkRunModel.R

@@ -6,7 +6,7 @@ lapply(df$model, function(model) {
     # Limit threshold for evolution of execution time (in %) between the 2 versions
     # Negative values of evolution are expected but we apply an error margin depending on execution time
     # decrease from 1.5 at 0.0ms to 0.5 at 10ms with a minimum at 0.5
-    threshold <- max(0.5, -0.1 * df[sel, 2] + 1.5)
+    threshold <- max(0.5, 2 - 0.1 * df[sel, 2])
     expect_lt(df$evolution[sel], threshold)
   })
 })