Parallelizing FORTRAN 77 codes is a non-trivial process. As parallel computers approach new levels of complexity and performance, the task of program conversion becomes almost impossible without the aid of a software tool like BERT.

BERT is designed to answer many of the difficult questions concerning your program. All to often, users will ask; "Can it run in parallel ?". Unfortunately, the answer to this question will not provide any insights into an efficient parallelization. The better questions is; "Should it run in parallel ?"

The following questions and answers provide an overview of BERT for the parallel computers.

Can my program run in parallel ?

The concept of parallel computing offers exciting possibilities. Accessing any new technology requires asking the right questions. Parallel computing is no different.

As with any new technology, there are often incorrect "notions" about what is possible and what is not possible. One such notion is that almost unlimited parallel computing is possible by simply connecting large amounts of processors. Ask any one who has used a parallel computer and you will find that this is just not true.

Another incorrect notion is "anything that can be run in parallel should be run in parallel". More often than not, a FORTRAN 77 program is converted for a message passing parallel environment, either by hand or with some tool, and the user is surprised to find the performance gain is not what is expected. Only after careful investigation and experimentation can the program be converted with good results.

So what is the problem ? Well, for one thing, some software tools and experienced people can do a pretty good job of determining concurrent portions of a program. Concurrent portions are those parts of a program that are independent.

The problem is whether a concurrent portion of the program should be executed on multiple processors. The simple question "Can it run in parallel ?" now becomes "Should it run in parallel ?"

Should my program run in parallel ?

If communication time is negligible compared to computation time, then all concurrent portions should execute in parallel. In reality, communication time is often not negligible. Because communication time plays an important role in parallel performance, the user is forced to ask the question:

"Should this concurrent portion be executed on multiple nodes ?"

The answer is not simple to determine. It requires knowledge of communication time, computation time, and the number of processors.

Conversion tools that ignore this question have solved half the problem. Making all concurrent parts of a program parallel may actually decrease performance. Of course, there are some obvious parallel FORTRAN codes, but other codes may not have obvious parallelism and even with "concurrency detectors" users must resort to a trial and error process to determine a good parallelization.

Finally, consider the problems of moving a parallel program from one platform to another. We all know parallel codes may be run on workstation clusters, but is a parallelization for workstations the best one for a dedicated parallel machine ? Or will a parallelization done for a parallel computer work better with additional nodes or an upgraded machine ?

These can be difficult and expensive questions to answer.

I do not think automatic tools create efficient codes.
Why should I use such a tool ?

There is no substitute for an experienced (and expensive) programmer, but there are tools that can help reduce the effort required for parallelizing FORTRAN programs. In fact, as you will learn, a tool like BERT can make determinations about your program that would be nearly impossible to do by hand.

If you have no experience with parallel programming, BERT can examine your program and produce a parallelization quickly. Further optimizations can be done using BERT as a guide.

BERT will leave your original source code unmodified (except for comments). Optimizations are done within the context of your original source code. You can make changes in your source program, analyze it with BERT, and determine if your change will lead to better parallelizations - without running on a parallel computer!

What does BERT do exactly?

BERT is designed to automatically convert FORTRAN 77 to message passing FORTRAN programs. It is a "source to source" conversion tool. How automatic the conversion is depends upon how the original program was written. There is no software tool that can do a fully automatic conversion for all FORTRAN programs.

Why is BERT different from other conversion tools?

The BERT analysis uses communication time and computation time to determine the best method for parallelizing your program. Other tools often do not consider the effect of these times. BERT will identify parallelism and it will only schedule parallelism if it is expected to produce a faster program. One thing to remember "concurrent does not mean parallel".

Because BERT includes communication time, it is possible to give a performance estimate for the conversion. With this information, it is possible to determine what parts of your program will give the best "parallel payoff".

What do you mean "concurrent does not mean parallel" ?

Using our definition, concurrent means that two or more things can be done independently, but not necessarily on different processors. Multiple users on a workstation constitute a concurrent operation. Parallel means that two or more things can be done independently at the same time on different processors. Just because something is concurrent does not mean that you can increase the program speed by making it parallel. BERT will examine your program and schedule those concurrent sections that can be economically executed in parallel.

Does communication time determine whether something that is concurrent is scheduled as parallel ?

Yes, but you must also consider the computation time. Because communication time and computation time vary from machine to machine, BERT will convert and schedule your program based on these parameters. This information is contained in a special file (fun.std) which is determined for each host parallel computer. Perhaps what is most interesting is that, depending upon communication and computation time, BERT will create different programs for different machines. "What is parallel for some may not be parallel for all."

Do you mean if I spend the time to convert a program by hand for a cluster of workstations using PVM and then move it to an parallel computer that it may not be efficient ?

Exactly! And, the converse is true; if you convert a program for a parallel computer, that conversion may not be best for another environment. It is a very difficult task to consider all the parameters that will determine a good parallelization. Because parallel environments are so different, it is almost an impossible task to do by hand. Furthermore, there are various parallelization models that one could consider that will also have an effect on performance.

What do you mean by parallelization models?

There are several ways to solve a problem in parallel. For instance, the most common "static" method is to spread a large array across several processors and perform independent computations on each processor. Depending on the program, there are other methods that may produce better performance if a "dynamic" model is used instead of a static mode. Again, which model is best depends upon communication time and computation time. More information about parallelization modelsis available.

Can BERT help me determine the best model?

Yes. BERT supports both a static/data parallel model and a dynamic/data flow model. Actually, version 1.1 of BERT will support multiple models within the same program.

What about standards? Good question. First, as a user of BERT, you will continue to interact with your FORTRAN program as you have in the past. Your original program remains intact. You or BERT may provide parallelization hints about your program, but these are only in the form of comments. By maintaining your original program as FORTRAN 77, BERT can decide the best conversion for the current environment.

For message passing, BERT currently supports PVM and some host specific libraries. Recall, however, that just because BERT has produced a PVM version of your program for one environment, this is no guarantee that this program will run optimally in a different environment. This is why a parallel program that is said to be "standard PVM" is misleading. It will run on several environments, but it may not run efficiently on all of these.

It seems that BERT is a good tool because the communication times may be quite different from machine to machine?

You are absolutely correct. BERT can provide insights into the best parallelization for a specific machine that may be difficult to determine otherwise. It is possible that a parallelization produced by BERT on one machine may not be efficient for another machine.

Why is BERT different from other FORTRAN compilers?

There are several differences. BERT is a "source-to-source" translation tool. Source code produced by BERT must be compiled by a FORTRAN compiler. The most important difference between BERT and FORTRAN compilers is that BERT does interprocedural analysis. This analysis allows automatic parallelization to be performed - including loops with function calls.

Another major difference is parallelization models. BERT currently supports a message passing environment and two parallelization models - dynamic and static. Depending on the application, one model may be more beneficial than the other. Having both models at your disposal allows the best parallelization to be determined for your application.

Is BERT available now ?

Yes, please see the down load page.