On 2/22/2011 12:05 PM, Patricia Shanahan wrote:
On 2/22/2011 12:16 AM, Peter Firmstone wrote:
Patricia Shanahan wrote:
I want to get going on some performance tuning, but believe it is best
guided and controlled by well-organized benchmarks. To that end, I
propose adding a place for benchmarks to the River structure.
We will need several categories of benchmark code:
1. System level benchmarks. These benchmarks measure some public
features, such as the outrigger JavaSpace implementation. For these, I
think a similar structure to QA may be best. However, I need to
understand how the QA harness links together clients and servers, and
whether it has any special performance implications. We may need, for
example, to add network delays to properly score implementations that
involve different amounts of communication.
2. Internal benchmarks. These are more like unit tests, and need to
mirror the main src package structure so that they can access
non-public code.
3. Experimental code. In some situations it is useful to do run-offs
between two or more implementations of the same class. We cannot have
two classes with the same fully qualified name at the same time, so
this type of test will need special copies of the classes with
modified class names or package names. In addition to actually doing
the tests and picking the implementation to go in the trunk, it is
useful to keep discarded candidates around. One of them may turn out
to be a better basis in a future performance campaign.
Thoughts? Alternatives? Comments?
Patricia
+1 to 1 and 2, not sure how to handle 3 - Peter.
I wonder if we could have a location for long term experimental code in
skunk?
If the experiment into a modular build is successful, (my apologies for
my recent lack of time), we could simply create an experimental module
and compare it against the original.
We won't always be able to integrate an experiment with its proper package until
after the experiment has been done.
For example, my recent FastList changes involved a change in how a FastList user
scans the list, from one based on list.head() and node.next() to making FastList
Iterable. I did not do the changes to the rest of outrigger to compile with the
new interface until after I had assured myself that at least one Iterable
implementation was as fast as the old implementation.
I'm also dubious about doing performance comparisons with different environments
for the code being compared. My ideal is a program that can cycle among
implementations in a single run. Next best is a program that measures a run-time
selected implementation, but with everything except the code under test
unchanged. Everything involved must be built with the same compiler version and
parameters, so I strongly prefer a single build.
I'm not sure how all these issues would be handled in the modular build
environment.
One thing you might do is use interfaces to bridge the gap. In particular, for
a class, you can move the class to a newly named class, and make the old class
name into an interface which is then implemented by the new test/experiment
class. This won't work across the board, so you might then have to use an
abstract class if there is a variable reference or other implementation detail
that requires a class instead of an interface.
Just some thoughts.
Gregg
