Revised benchmark procedure.

Now we take the difference of the time to process the input
and the time to run with no input.  This compensates for
slow startup time in dynamic languages.

See comments on 2dcef8a

Makefile

@@ -127,6 +127,7 @@ progit/progit.md: progit
 bench: $(BENCHFILE)
 	{ sudo renice 99 $$$$; \
 	  for x in `seq 1 $(NUMRUNS)` ; do \
+	  /usr/bin/env time -p ${PROG} </dev/null >/dev/null ; \
 	  /usr/bin/env time -p ${PROG} $< >/dev/null ; \
 		  done \
 	} 2>&1  | grep 'real' | awk '{print $$2}' | python3 'bench/stats.py'

bench/stats.py

@@ -3,7 +3,15 @@
 import sys
 import statistics
 
-values = [ float(x) for x in sys.stdin.readlines()]
+def pairs(l, n):
+        return zip(*[l[i::n] for i in range(n)])
+
+# data comes in pairs:
+#    n - time for running the program with no input
+#    m - time for running it with the benchmark input
+# we measure (m - n)
+
+values = [ float(y) - float(x) for (x,y) in pairs(sys.stdin.readlines(),2)]
 
 print("mean = %.4f, median = %.4f, stdev = %.4f" %
     (statistics.mean(values), statistics.median(values),

benchmarks.md

@@ -4,23 +4,30 @@ Some benchmarks, run on an ancient Thinkpad running Intel Core 2 Duo at 2GHz.
 
 |Implementation     |  Time (sec)| Factor  |
 |-------------------|-----------:|--------:|
-| Markdown.pl       | 2921.30    | 14606.5 |
-| kramdown          | 20.76      |   103.8 |
-| PHP markdown      | 20.475     |   102.4 |
-| lunamark          | 6.195      |    40.0 |
-| cheapskate        | 5.645      |    28.2 |
-| peg-markdown      | 5.330      |    26.7 |
-| **commonmark.js** | 2.730      |    13.6 |
-| marked            | 1.870      |     9.4 |
-| discount          | 1.660      |     8.3 |
-| **cmark**         | 0.320      |     1.6 |
+| Markdown.pl       | 2921.24    | 14606.2 |
+| PHP markdown      | 20.85      |   104.3 |
+| kramdown          | 20.83      |   104.1 |
+| lunamark          | 6.295      |    31.5 |
+| cheapskate        | 5.760      |    28.8 |
+| peg-markdown      | 5.450      |    27.3 |
+| **commonmark.js** | 2.675      |    13.4 |
+| marked            | 1.855      |     9.3 |
+| discount          | 1.705      |     8.5 |
+| **cmark**         | 0.315      |     1.6 |
 | sundown           | 0.200      |     1.0 |
 
+
 To run these benchmarks, use `make bench PROG=/path/to/program`.
 
 The input text is a 10MB Markdown file built by concatenating 20 copies
 of the Markdown source of the first edition of [*Pro
 Git*](https://github.com/progit/progit/tree/master/en) by Scott Chacon.
 
-`time` is used to measure execution speed.
+`time` is used to measure execution speed.  The reported
+time is the *difference* between the time to run the program
+with the benchmark input and the time to run it with no input.
+(This procedure ensures that implementation dynamic languages are
+not prenalized by startup time.) A median of ten runs is taken.  The
+process is reniced to a high priority so that the system doesn't
+interrupt runs.