add comment on benchmarking approach

Signed-off-by: martinvuyk <[email protected]>

stdlib/benchmarks/collections/bench_string.mojo

@@ -223,8 +223,17 @@ fn bench_string_is_valid_utf8[
 # ===----------------------------------------------------------------------===#
 def main():
     seed()
-    var m = Bench(BenchConfig(num_repetitions=1))
-    m.bench_function[bench_string_init](BenchId("bench_string_init"))
+    var m = Bench(BenchConfig(num_repetitions=5))
+    # NOTE: A proper way to run a benchmark like this is:
+    # 1. Run the benchmark on nightly branch with num_repetitions=5 and take the
+    #    **median** value for each function, length, and language that is to be
+    #    measured.
+    # 2. Then run the benchmark on num_repetitions=1 if you want faster results
+    #    during development of your branch.
+    # 3. When ready to make statements about speed improvements, first run the
+    #    benchmark again with num_repetitions=5 and take the **median** of that.
+    # 4. Make a table and report the new **median** numbers and the markdown
+    #    percentage improvement over nightly version (new - nightly)/nightly.
     alias filenames = (
         "UN_charter_EN",
         "UN_charter_ES",
@@ -236,6 +245,8 @@ def main():
     alias new_chars = ("A", "Ó", "ل", "И", "一")
     alias lengths = (10, 30, 50, 100, 1000, 10_000, 100_000, 1_000_000)
 
+    m.bench_function[bench_string_init](BenchId("bench_string_init"))
+
     @parameter
     for i in range(len(lengths)):
         alias length = lengths.get[i, Int]()