forums.silverfrost.com

[Wilfried Linder]

Hi (and sorry for my question - up to now I really have no idea about it):

nowadys more and more PCs have more than one processor. What is the way to use all of them? Is it necessary to modify the code in the way that the programme subdivide the work, or does the operating system manage this? In other words: If I use my "standard" fortran code, will only one single processor run the programme?

Thanks in advance and best regards,
Wilfried

[JohnHorspool]

[Wilfried Linder]

hmmm... but first thank you!

Part of my work is image processing what simply means that operations must be done on millions of pixels. For instance (to use a very simple example) imagine that for each pixel I would like to carry out the function new_value = 2*old_value. Also imagine that the image size is 10000 by 10000 pixels. If I have 4 processors I'ld like to handle the work like processor 1 = lines 1 ... 2500, processor 2 = lines 2501 ... 5000 and so on.

As far as I understood, first I must get information about the number of processors (how to do this?). After that I must split the work using (say) four times start_pprocess@, then waiting until the last processor has done the work, then matching the results together to a new image. Is that right?

Again thanks in advance...
Wilfried

[JohnHorspool]

Wilfried,

[Wilfried Linder]

Thank you very much!

Sorry for my questions... May be I need more precise information about ftn95. I only have the manuals salfordftn95.pdf and clearwin.pdf. For instance, the fortran manual tells me about a function getenv@ but didn't show any list of possible valid parameters (like 'number_of_processors' you wrote).

Also, within these two documents no information is given about start_pprocess@. I don't what to disturb you too much - so, can you please give me an address or so to get more actual manuals? It will be the best that I first read them carefully before going on.

Best regards,
Wilfried

[PaulLaidler]

start_pprocess@ is described in the current ftn95.chm under start_process@.

[JohnHorspool]

Wilfried,

You will not find a list of arguments for getenv@ in the manuals, it is not necessary.


Instead go to:-

Start --> Settings --> Control Panel --> System Icon --> Advanced tab --> Environment variables

and you will see a list of them with their current values, gentenv@ can retrieve the value of any of them.

If that has whetted your appetite then you can get even more information about your OS and hardware using the function REGOPENKEYEX to extract stuff from the registry (but thats more involved than using getenv@ !)

[Wilfried Linder]

Thanks again!

Paul: I tried to get the ftn95.chm from your website (Documentation > Silverfrost FTN95 User Guide), but in vain. I tried it in my office, I tried it at home, I tried to read and also to download, but all times an error occurs. I really have no idea what happened...

John: I found the variables and it works. Thanks!

Best regards,
Wilfried

[LitusSaxonicum]

Not all that long ago I posted all the code necessary to get a whole bunch of processes running. You also have to make sure that they run on separate cpu cores. You can do that too. The facilities to get all the separate applications to talk to each other are standard things in Clearwin. I thought that no-one was interested!

Search for "Home made multithreading"

Eddie

[Wilfried Linder]

Hi friend from the saxonian coast, THAT'S WHAT I NEED;-) Thanks, I'm very interested and I will report about my work - and disturb you with questions if necessary ...

Wilfried

[Wilfried Linder]

Very interesting, indeed... I made the first tests and I hope that I begin to understand the principles.

As far as I see up to now, I must split my programme into several executables, right? Or is there any alternative to devide the work to several processors from within a single executable? What I mean is something like this:

[LitusSaxonicum]

Hi Wilfried,

In the Dawn of Time, Salford FTN77 supported multithreading. Later, following sale of soul to devil (which happened in 2 stages, and which is what I call the move to Windows!), multithreading is not officially supported with Clearwin. Whether or not it is with .NET, I can't tell you.

The principles involved in running your Clearwin application on a multiprocessor machine is that up to a point, Windows does appear to use all the cores, but in a way that gives you no real nett benefit - although you can run multiple applications to raise the load on each cpu core.

Each window is a sort of independent application - again, up to a point. I could find no way of making particular windows attach to particular cpu cores, or even to different ones, although some sort of allocation and scheduling must be done by windows.

What I was able to do was to create separate standalone applications that could talk to each other. Then, I found that I could associate particular applications with a given cpu core (through task manager). I imagine that one could also create a flurry of mini-applications that would cause Windows to associate some with one core, and others with another core. Indeed, you can obtain (buy!) software that makes the association of a program with a particular core "persistent" (I wrote a post in reply to a "core duo" discussion thread on this).

In your example, you wouldn't get any benefit, because you wouldn't "switch_to_processor@(2)" until processor 1 had completed the tasks set in "tell_him_what_to_do@" - whereas if you start_pprocess@, you do move to the next statement (you don't with "start_process@"). In order to do what you want within a single application, there needs to be specific facilities for multi-threading, and they don't exist. However, a set of intercommunicating applications are, by definition, multithreaded, and Clearwin does have that.

In a multithreaded system, you have to be able to handle the results coming back in an order that is different from the tasks being issued. The phrase that springs to mind is "client-server computing". As that is american, the "client" is a customer in a restaurant, the "server" is a waiter. If the client has only one server, then the dishes ordered are likely to be served in the correct order. However, if there are several servers, and one is sent for starter, one for main course, and one for wine etc, then the dishes could be delivered in any sequence!

This analogy is nested, for example, the server is effectively the client for the chefs in the kitchen.

With conventional Fortran, there is only one server, and you don't ask him for the main course until after the starter has been delivered!

I think you need a different mindset to cope with multithreading, whichever way you do it.

Eddie

[Wilfried Linder]

Thank you, Eddie.

You wrote "Later, following sale of soul to devil (... which is what I call the move to Windows!), ... ". Really nice to read that other people have the same feeling about that like me. But of course, we will not stop the time

Wilfried

[LitusSaxonicum]

Hi Wilfried,

That was just a throwaway expression. If it wasn't Windows, it would be GEM, or possibly some other Mac lookalike. FTN77 with DBOS wasn't just the fastest and "bestest" at compiling, it also had multithreading (although why, I never worked out), excellent graphics (up to VGA), loads of extra facilities like virtual memory AND some of the fastest run times for the compiled code.

I don't think that the Windows way of working is always the best. It certainly doesn't fit easily with Fortran. Quite obviously, Salford/Silverfrost did a fantastic job with Clearwin, but there are some loose ends.

The tradeoff for getting Clearwin is that you no longer have the fastest EXE. If that bothers you, write the front-end in FTN95/Clearwin and the back-end in (say) Intel Fortran. Run the back-end using start_pprocess@ ...

Regards

Eddie

[JohnHorspool]

Eddie,

As I recall the graphics under DOS and FTN77 was only limited by the graphics card you had installed. I recall getting SVGA and beyond with FTN77 and certainly a lot better than Win3.1 could achieve!

cheers
John