Re: [Caml-list] Re: SysThreads and DLL's

["David McClain" <barabh@qwest.net>]

After another day spent experimenting, I have a wonderfully working system
with OCaml now. I spawn a thread to perform OCaml initialization and then
sit in an event loop looking for request messages. These messages, sent with
PostThreadMessage() from any old thread to the OCaml event loop thread, are
simply forwarded to the OCaml DLL code. The message carries a tag indicating
which DLL routine to run, the wParam carries the ThreadID of the calling
thread. The lParam of the message carries a pointer to the actual C
arguments, and this is bundled up as an opaque type for OCaml. Inside the
DLL OCaml unmarshalls those parameters as need be.

This event loop needs to have the GetMessage() call surrounded by
enter_blocking_section() and leave_blocking_section(), as I rediscovered
last night. This isn't mentioned in the documentation, but there are a few
comments in the OCaml sources. Thanks so much for providing these sources!!
Those comments indicate, as one might expect, that only one Windows thread
at a time is permitted to churn about inside OCaml. This is achieved by
means of a Mutex object used backwards... enter_blocking_section() releases
this Mutex, and leave_blocking_section waits until the Mutex can be grabbed
again. Very cute!

Inside the DLL, OCaml can keep a stable full of threads, as it wishes, and
the incoming DLL requests can be spawned off on as many threads as you like.
After each DLL request is finished working, they need to send back a simple
reply to the waiting threads. This is done by passing an HRESULT via another
PostThreadMessage() call to the calling thread's ID as passed along in the
original request message.

The result really churns along and appears to satisfy all my needs for
interfacing with arbitrary foreign threads.

Not very messy at all, actually. The startup of a special initialization and
marshalling thread only takes a few lines of C (maybe 20 in all). The
protocol of sending messages and waiting for a response is maybe another 20
lines of C.

OCaml is sheer pleasure to program!!

Cheers,

- David McClain, Sr. Scientist, Raytheon Systems Co., Tucson, AZ

----- Original Message -----
From: "David McClain" <barabh@qwest.net>
To: <caml-list@inria.fr>
Sent: Saturday, December 01, 2001 2:57 AM
Subject: [Caml-list] Re: SysThreads and DLL's


> Okay, after some experimenting here is what I found out about OCaml and
> LispWorks Lisp...
>
> 1. I set up a test in both languages that executes a nearly infinite loop
to
> generate massive amounts of garbage -- hence to force many GC operations.
> Threading is not explicitly used in either of these two programs, although
> the OCaml program was compiled with the -thread option and linked with the
> native thread libraries. However, since I never initialize the threading
> package (I have them initialize only when used with the Lazy.force trick),
> the threading engine of OCaml should not be running.
>
> 2. I packaged these two programs up as DLL's for Win/NT 4.0
>
> 3. I wrote a simple C driver program to link to these DLL's, initialize
> them, and the spawn a raft of threads to exercise that infinite loop all
at
> the same time....
>
> 4. LispWorks Lisp survives this test indefinitely. Unfortunately, a
positive
> result here is actually indeterminate. All I can say is that it hasn't
blown
> yet....
>
> 5. OCaml survives briefly and then gives GC failures. The test is
> nondeterministic, and so the failures occur at varying durations of the
> test, but OCaml always shows early failures, relative to the Lisp.
>
> 6. Realizing that the OCaml GC must make use of thread specific
information,
> and hence, rendering this test invalid, I set up an umbilical chord
between
> incoming thread requests and one thread started up explicitly to handle
all
> OCaml DLL requests. This is simply two event loops, one on a thread
started
> specially to initialize OCaml and perform all calls into the OCaml code,
and
> the other is used by the calling thread to await a result from the OCaml
> thread. DLL calls are relayed to the OCaml thread by means of
PostMessage().
>
> 7. Doing this works great! BUT!!! Only one thread at a time can exercise
the
> OCaml code. And if the code is an infinite loop, none of the other threads
> get a chance to run.
>
> 8. I suppose that I could have OCaml spawn a bunch of worker threads and
> have them ready and waiting for assignments as they come in. I seem to
> remember the X-kernel doing something similar for marshalling network
> packets up and down the protocol stack. The idea would be to never allow
> these threads to die. When they finish an assignment they go back to the
> stable to await the next assignment. So that will be my next experiment.
>
> I find it interesting that LispWorks is able to survive this kind of
> punishment -- at least for as long as the few minutes duration that I
> watched. But, again, this is an indeterminate kind of answer. At least
with
> OCaml I *know* where I stand!
>
> BTW, the whole point of this exercise is to find a way to seriously shrink
> an oppressively burdensome code base written in frightful C/C++ with
> STL/ATL/COM/OLE and a bunch of other junk. I have already seen that both
> OCaml and Lisp produce roughly comparable shrinkages with improved
> capabilities to boot. The shrinkage is roughly 6 to 1.
>
> My test case is a COM/OLE server for translating network system wide
> filenames using recursive prefix mappings like Ousterhout used in the
Sprite
> OS. The C/C++ code is roughly 3 KLOC that I wrote several years ago, plus
a
> host of machine generated code to go with it. The Lisp and OCaml versions
> are roughly 600 LOC.
>
> Interestingly, OCaml is terser than Common Lisp, despite the lack of
> built-in richness that Lisp posesses. The syntax of Lisp is just more
wordy.
> So despite the lack of built-in library routines in OCaml, which I had to
> develop by hand, the overall result is roughly the same size between them.
> That even included writing an OCaml version of a Lisp-like PATHNAME module
> to parse and represent generalized pathnames. In a sense, OCaml is
> comparable to typical Scheme languages when it comes to built-in functions
> (probably better!).
>
> My experiments with LispWorks have been fruitful, but I have reservations
> about using it because of the extreme lack of documentation, and a
> demonstrated failure to be robust against punishment. Despite the results
of
> the test described above, whenever I generate the DLL for the pathname
> translations, it always fails to run properly the first time. I have to
kill
> the process and start up again. Thereafter it runs reliably as far as I
can
> tell.
>
> But with the OCaml version, it runs first time, every time. Like a real
> champ! The only drawback is the need for greater amounts of C wrapper code
> to make it into a DLL, and the need to produce a Makefile.
>
> My aim is code base reduction, and ease of generation, because I want to
> unload about 300 KLOC onto some unsuspecting maintenance programmer. I
need
> to get out from under this code burden!!!! So the jury is still out on
which
> way to go... Code base reduction is the same, ease of use is somewhat
better
> in Lisp, but robustness is significantly better in OCaml... It also helps
to
> have the source code!!
>
> Cheers,
>
> - D.McClain, Sr. Scientist, Raytheon Systems Co., Tucson, AZ.
> -------------------
> Bug reports: http://caml.inria.fr/bin/caml-bugs  FAQ:
http://caml.inria.fr/FAQ/
> To unsubscribe, mail caml-list-request@inria.fr  Archives:
http://caml.inria.fr
-------------------
Bug reports: http://caml.inria.fr/bin/caml-bugs  FAQ: http://caml.inria.fr/FAQ/
To unsubscribe, mail caml-list-request@inria.fr  Archives: http://caml.inria.fr