Skip to topic | Skip to bottom


Home

EelcoVisser
EelcoVisser.TransformationAspectsr1.1 - 27 Nov 2001 - 09:11 - EelcoVisser

Start of topic | Skip to actions
Ideas on the subject by Hedzer Westra.

Considerations for a Stratego Reporting Project

It can be very valuable to get meta-information about the execution of a transformation. This meta-information tells us something about how the input was transformed to the output. We discern two types of meta-data:

  • profiling information: how fast, what parts of the program are used the most,
  • effects: report the changes that were made, statistics about the data

Profiling

All profiling should be done automatically when and if the user requests this. One can think of extending the \texttt{iowrap} library function with a few options.

Profiling meta-data can be:

  • timing: read file, execute, write file (absolute and relative to total)

  • timing per strategy: time spent in each strategy, with and without time spent in called strategies. (==child coverage)

  • strategy counts: the number of times a strategy is called

  • maximum call depth

  • data information: tree size, number of nodes, number of terms thatuse a certain constructor, tree depth. Since this information is dynamic, a report should be made on a certain interval, or min/avg/max/stddev information should be calculated. It should be possible for the user to request the current data information using a SSL strategy.

Effects

Effects are application-specific and can therefore not be automatically provided by the compiler. The SSL can provide a standard mechanism to administer effects, though.

One can think of SSL support for counters, which the user can increase of decrease when certain events happen, e.g. a 'rule' succeeds. With a 'rule' we mean a transformation that has higher granularity or stricter conditions than a strategy, and therefore can't be automatically counted by the builtin strategy counting facility.

Another effect that one would want to measure is a user-defined statistic, e.g. the number of GO statements in the program that is being transformed. Calculating this statictic is of course the responsibility of the user. What the SSL could supply is strategies for doing statictical operations (min/avg/max/stddev/et cetera) on this data.

Creating and handling reports

The reports that are gathered during execution should be written to a standard Stratego report file. If would be very practical to extend iowrap with an extra option, so the user can specify where this file should be placed. A grammar should be defined in the GB (and more importantly: a Stratego signature), so standard tools can be developed to handle reports from all kinds of program executions.

The following report handling tools could be developed:

  • pretty printing (human readable reports)
  • GNUplot input generation (data and GNUplot definition)
  • LaTeX? tabular creation

It should be possible to combine reports to calculate statistics about e.g. how input size relates to exection time, generating plots for sets of transformed files, et cetera.

Combining reports should be done on two axes: transformation phase and transformed file. The transformation phase axis is about parsing, imploding, transforming and pretty printing one file. The transformed file axis is about sets of files being transformed.

The previous paragraph suggests that each component of a transformation system should generate reports. Since most XT tools are coded in Stratego, this should be fairly easy. Sglr is an important exception. It should be extended with report facilities. Another problem is the use of standard Unix tools. Maybe creating a wrapper program that generates a simplistic report can be considered.

A nice feature would be to compare output files or report files and generate reports about the differences.

Reporting so far

In the CobolX? project, a lot of the reporting that is described above is done, although it is done using mostly standard Unix tools like time, awk, diff and grep. A uniform reporting language is not defined. We use parsing of the output that our Stratego programs generate on stderr by calling debug, which is certainly not optimal. A lot of report grammars are handwritten (and rewritten often) to parse the textual program output.

The statistic calculations and profiling that is described above isn't done at all in CobolX?, with the exception of very crude timings using the Unix \texttt{time} tool in combination with AutoMake?. This gives a granularity at Unix process level, which is fine for generating plots that show what amount of time is spent on e.g. parsing. It doesn't tell us anything about where the transformation spends most of its time, though.


You are here: EelcoVisser > TransformationAspects

to top

Copyright © 2003-2017, Software Engineering Research Group, Delft University of Technology, The Netherlands