Find file
Fetching contributors…
Cannot retrieve contributors at this time
93 lines (79 sloc) 4.77 KB

Architecture of mod_parrot

Now that I've written 'architecture', I realise 'anatomy' sounds much cooler. But one can not have all. The design I had in mind is based upon a number of assumptions listed here:

  • Interpreter instantiation is expensive
  • The parrot interpreter is not thread-safe (the library as a whole is, though)
  • The parrot bytecode is not a stable format
  • The compiler API is relatively stable
  • Compilation is slow
  • The parrot C api is clunky (and relatively slow)
  • It is relatively tricky to get data out of C structs from within parrot
  • Buffering and copying the output of a script is expensive and redundant
  • It is not our job to parse query parameters or POST data
  • But we should report errors and log data

The first two assumptions mean that we would like a thread-local interpreter instance. I am not aware of the 'right' way to do this in the apache web server, but I would be rather surprised if there was none. For the time being, I can assume that workers are single-threaded processes, as I am developing on linux, and the prefork multiprocess module is common there (which is single-threaded per process). (Also, some people - in some enviroments - might actually want to run using a new instance every request. We should not punish them for this, and it should be an option).

Assumption 3 means that I do not expect people to precompile their programs before putting them online (as they do with ASP.NET and probably some other frameworks). Some people will no doubt distribute or serve their programs as PBC files (and mod_parrot should support this workflow) but in all likelihood most users will upload their scripts as source code. Also, parrot iterates rapidly and the bytecode format may change. However, as a temporary format, I expect the same script to be executed faster when stored in PBC format than in the original source format, as compiling is slow. (This is an excellent assumption to test by benchmark, by the way).

Assumption 4 is crucial to the implementation of so-called bootstrapping scripts, which are programs written in winxed with the goal of executing a given script within a certain enviroment. Using different bootstrapping scripts, mod_parrot could provide different enviroments for web applications to run in. The two most obvious examples are PSGI and CGI, which have similar but not equal execution enviroments. Of course, different bootstrapping scripts will probably share a lot of code.

The fifth assumption means that it would be wise - in many cases - to cache compiled bytecode files so as to skip compilation. Note that this does not mean we can skip loading the language enviroment, as this may still be necessary.

Assumption six and seven (as well as nine, in a way) concern the balance between doing much in C versus doing much from within parrot, using winxed. The goal is to keep each job as simple as possible, so the idea was to convert the relevant parts of the request_rec structure into a hash which is passed - alongside the apache options, did I mention apache module options? They will be stored in a hash and passed as well - to the bootstrapping script. The bootstrapping script can then setup the enviroment. For the most part, this hash will contain a simple string->string mapping. However, this hash will also contain the request_rec structure itself, since it is a filehandle of sorts. More on that below.

The eight assumption concerns output. As of apache 2.0, it is relatively tricky to get the actual on-the-wire socket because of the filtering API. Thus, output is buffered on the server side in 'buckets'. The good thing is that this is relatively transparent to the user if that user is a C programmer and cares to call the right functions. These functions take the request_rec structure in the place of a regular filehandle and are call ap_puts and ap_write or similar. The best solution AFAIK is to replace parrots standard IO handles with a PMC that calls these functions, using the request_rec structure. I will have to discuss this with whiteknight. If this method should fail, it is still possible to return from the bootstrapping script to the C module the output as a string (or series of strings, because headers need to be handled separately). This would mean buffering output thrice (from parrot, to mod_parrot, to apache), which is memory-inefficient as well as slow, and prohibits HTTP streaming.

The 10th and last assumption means that mod_parrot needs to report errors to the apache error log. The level of detail desired may be configured at the module level. Within the parrot interpreter, standard error will do nicely for this purpose.


All source code goes to src/ The 'module' (C code) goes to src/module The loaders (winxed) go to src/loader The library for loaders go to src/library

Test to go t/