perf: call _generatechangelog() instead of group()
authorGregory Szorc <gregory.szorc@gmail.com>
Mon, 06 Aug 2018 10:43:05 -0700
changeset 38977 a1f694779b2f
parent 38976 f7228c907ef4
child 38978 d2ab0db89465
perf: call _generatechangelog() instead of group() Now that we have a separate function for generating just the changelog bits, the perf command should call it so it gets more accurate behavior. This changes the results of this command on my hg repo significantly: ! wall 1.390502 comb 1.390000 user 1.370000 sys 0.020000 (best of 8) ! wall 1.768750 comb 1.760000 user 1.760000 sys 0.000000 (best of 6) Profiling seems to reveal that ~20% of execution time is spent in progress bar accounting and printing! If we run with progress.disable=true: ! wall 1.639134 comb 1.650000 user 1.630000 sys 0.020000 (best of 7) A nice speedup. But profiling still shows a good chunk of time being spent in progress bar accounting code. The reason is that the progress bar is conditionally enabled via an argument to cgpacker.group(). The previous code in perf.py calling into group() did not enable the progress bar but _generatechangelog() always does. I think it is important for the perf* commands to capture real-world use cases. And this code always runs with an active progress bar. So the regression is acceptable. That being said, terminal printing performance can vary substantially. I don't think perf* commands should test terminal printing unless explicitly desired. So I've disabled progress bar printing in this command. Differential Revision: https://phab.mercurial-scm.org/D4134
contrib/perf.py
--- a/contrib/perf.py	Fri Aug 03 17:59:56 2018 -0700
+++ b/contrib/perf.py	Mon Aug 06 10:43:05 2018 -0700
@@ -663,21 +663,20 @@
     By default, all revisions are added to the changegroup.
     """
     cl = repo.changelog
-    revs = [cl.lookup(r) for r in repo.revs(rev or 'all()')]
+    nodes = [cl.lookup(r) for r in repo.revs(rev or 'all()')]
     bundler = changegroup.getbundler(version, repo)
 
-    def lookup(node):
-        # The real bundler reads the revision in order to access the
-        # manifest node and files list. Do that here.
-        cl.read(node)
-        return node
-
     def d():
-        for chunk in bundler.group(revs, cl, lookup):
+        state, chunks = bundler._generatechangelog(cl, nodes)
+        for chunk in chunks:
             pass
 
     timer, fm = gettimer(ui, opts)
-    timer(d)
+
+    # Terminal printing can interfere with timing. So disable it.
+    with ui.configoverride({('progress', 'disable'): True}):
+        timer(d)
+
     fm.end()
 
 @command('perfdirs', formatteropts)