[Python-Dev] explanations for more pybench slowdowns

[Jeremy Hylton]

We looked at the SecondImport test case today.  It's a good test case
for programs that execute "import os" in a time-critical inner loop
:-).

The primary reason it is slower is the import lock that was added
after 1.5.2.  The benchmark, run in isolation, spends about 6 percent
of its time in the locking code.  Since it only spends about 20
percent of its time actually doing imports, this is a pretty
substantial cost.

It seems possible to eliminate some of the cost by using a special
marker in sys.modules that means: "This is not a module, but it's
being loaded by another thread."  But Guido doesn't sound interested
in optimizing programs with imports in inner loops.

Jeremy