[pypy-commit] extradoc extradoc: write abstract

fijal noreply at buildbot.pypy.org
Sun Mar 18 17:01:00 CET 2012

Author: Maciej Fijalkowski <fijall at gmail.com>
Branch: extradoc
Changeset: r4152:cc09a42ef4f7
Date: 2012-03-18 18:00 +0200

Log:	write abstract

diff --git a/talk/ep2012/tools/abstract.rst b/talk/ep2012/tools/abstract.rst
new file mode 100644
--- /dev/null
+++ b/talk/ep2012/tools/abstract.rst
@@ -0,0 +1,19 @@
+Performance analysis tools for JITted VMs
+When writing code in C, the mapping between written code and compiled assembler
+is generally pretty well understood. Compilers employ tricks, but barring
+few extreme examples they're, typically within 20% performance difference
+from "naive" compilation. Tools used to asses performance are usually
+per-function profilers, like valgrind, oprofile or gnuprof. They all can
+display time spent per function as well as cumultative times and call
+chains. This works very well for small-to-medium programs, however large,
+already profiled programs tend to have flat performance profile with unclear
+message from profilers.
+To make matters even more complicated, in JITted VMs, like PyPy, mapping
+between assembler and high level language, like python, is very unclear and
+not well known. Writing "good" vs "bad" python (from the JIT perspective),
+can make a 20x performance difference. I would like to show current profilers,
+their use cases and shortcomings as well as make the case that we need much
+better tools to deal with the current situation.

More information about the pypy-commit mailing list