COBOL performance for Complex arithmetic expressions

delago · New User Joined: 29 Jul 2005 Posts: 21 Location: Brazil

Hi gu'ys.

I need to talk about a topic, because I need sugestions for a future project.

The sitution: a project where a need to calculate a risc of financial operation generic. In the model especificated for the user (my client), I will use a lot of arithmetical expressions with hard difficult level.

Example a arithmetical express:
VaR_i^cred=max{max{K_i;piso_i_EAD_LGD}×LGD_i;piso_i_EAD }×EAD_i+SD_i×LGD_i

In the example, a need to run this arithmetical expression any times, because have a payment's month-a-month. Any times I using the last results for arithmetic expressions futures.

The concept is very hard to know.

With programing with language cobol, the codification is the bigest and hard to know. And the worst: the performance is not better the user (my client) will waiting on the result.

My question: what language and enviroment I will use to better results for cmplex arithmetical expressions?

I thank any suggestions.

Tks.

Fernando Delago - Brazil/SP

Bill Woodger · Posted: Thu Apr 07, 2011 6:07 pm

Are you saying you don't want to use Cobol? So why ask the question in a Cobol forum?

Phrzby Phil · Posted: Thu Apr 07, 2011 6:09 pm

How many millions of these calcs are you doing per day?

prino · Posted: Thu Apr 07, 2011 6:11 pm

Every mainframe compiler worth its price will (should) result in (more or less) the same code. If you want to speed up tings, you will have to move invariant expressions out of loops and possibly cache intermediate results.

Bill O'Boyle · Posted: Thu Apr 07, 2011 6:49 pm

Language Environment offers Callable Service routines for a wide variety of mathematical calculations. COBOL/370 (and greater) offers comparable FUNCTION's as well, but not to the extent of LE.

LE can be used t/w most languages, including Assembler, but the version/release (for HLL) must be compatible.

Bill

Robert Sample · Posted: Thu Apr 07, 2011 6:51 pm

COBOL is perfectly adequate at complex arithmetic. Things like USAGE and number of digits (before AND after the decimal point) can impact the calculations but should not stop you from using COBOL.

FORTRAN was designed for mathematical calculations. I haven't used it in probably 18 - 20 years but it might be worth considering.

In general, though, as prino mentioned, the compilers all generate close to the same code so any optimization achieved has to be done by hand using the methods described in his post and numerical analysis books. And don't expect magnificent savings -- once you've got a baseline time for execution, you're only likely to get tincremental benefits no matter what you do because sometimes things just take a certain amount of time, period.

PeterHolland · Posted: Thu Apr 07, 2011 6:52 pm

PL/1, Fortran, APL?

Kjeld · Posted: Fri Apr 08, 2011 5:02 am

All compiled mainframe languages basically produce the same code for arithmetic expressions.
The important thing is to define the operands in the compute statement as binary in order to avoid any type conversion packed or display formats during the computation. Intermediate results are ususally stored in the registers for fastest computations.

Newer z-series have special floation point instructions, that alledgedly can be used by some languages. PL/1, Cs and Enterprise Java have been mentioned.

I assume that you would be in total control of the computation if you use Assembler.

Bill O'Boyle · Posted: Fri Apr 08, 2011 6:00 am

LE Mathematical Callable Service routines, V1R12.0, Language Environment Programming Reference -

publibz.boulder.ibm.com/cgi-bin/bookmgr_OS390/BOOKS/CEEA31B0/2.1.8?DT=20100625181724

Examples -

publibz.boulder.ibm.com/cgi-bin/bookmgr_OS390/BOOKS/ceea31b0/2.4.4?ACTION=MATCHES&REQUEST=arithmetic&TYPE=FUZZY&SHELF=&DT=20100625181724&CASE=&searchTopic=TOPIC&searchText=TEXT&searchIndex=INDEX&rank=RANK&ScrollTOP=FIRSTHIT#FIRSTHIT

Sample calls to math services -

publibz.boulder.ibm.com/cgi-bin/bookmgr_OS390/BOOKS/ceea31b0/2.4.2?ACTION=MATCHES&REQUEST=arithmetic&TYPE=FUZZY&SHELF=&DT=20100625181724&CASE=&searchTopic=TOPIC&searchText=TEXT&searchIndex=INDEX&rank=RANK&ScrollTOP=FIRSTHIT#FIRSTHIT

Bill

delago · New User Joined: 29 Jul 2005 Posts: 21 Location: Brazil

Hi folks,

Clarifying some doubts:

- My process in my job increases exponentially with each execution, for dealing with complex arithmetic calculations for each tranche of a financing transaction. Example: In an operation of 60 services will be implemented over 50 calculations using values up to six decimal places, and this process can be performed with supplies 60 to 10 times (ie a prediction).
- In the first measurement of the performance of online transaction, we got almost 7 seconds of the end time. After some changes (almost restructuring of the system) can reduce the time by almost 1 second.
- I'm trying to prevent future re-implementations to not degrade performance.
- Some time ago, I heard that some programming languages to lower level, had better performance on the mainframe. This statement is true?

Thanks guys.

Fernando Delago - Brazil/São Paulo

Bill O'Boyle · Posted: Fri Apr 15, 2011 1:00 am

Are you using Floating Point (COMP-1/COMP-2) or Fixed Point (COMP-3) data types?

Floating Point in COBOL can be very costly, as it may cause the compiler to BALR/CALL run-time routines.

Fixed-Point (COMP-3) can support up the 16-Bytes, if your compiler supports the ARITH(EXTEND) option (OS/390 COBOL 2.2.1 and greater).

Could you post some of your calculations, without causing any trouble with management, as it's hard to visualise what you're doing.

COBOL does come with some extra "baggage" and Assembler may perform these calculations more efficiently.

If your issuing (for example) WS-AMOUNT ** 6 (WS-AMOUNT to the power of 6), then you're performing exponentiation on fixed-point data, by virtue of the "**".

Bill

prino · Posted: Fri Apr 15, 2011 1:17 am

Can you just show us a few of these calculations? A calculation is a calculation, and not a trade-secret. It might give us a much better understanding of what's required.

delago · New User Joined: 29 Jul 2005 Posts: 21 Location: Brazil

Ok folks... it's a few...I "break" in little arithmetic expressions for better performance, because read in IBM manual this alteration help me.

Phrzby Phil · Posted: Tue Apr 19, 2011 12:10 am

I repeat my question from April 7:

Bill O'Boyle · Posted: Tue Apr 19, 2011 12:24 am

Is there a good reason why you're using COMP-2 (Long Floating Point)?

Besides being a "Challenging" format for COBOL, floating point may not yeld the correct expected results and truncation can take place within intermediate results.

If COMP-2 is your desired final format, then perform the calculations in fixed-point, either packed-decimal or binary and MOVE the result to the COMP-2 field when you're done.

Compile your code with LIST,NOOFFSET compile options (generates Assembler expansion) and you might be surprised that some of these calculations require a CALL/BALR to a COBOL run-time routine, which you don't want unless you don't have any other choice.

Bill

Bill Woodger · Posted: Tue Apr 19, 2011 6:22 am

Locate and digest "IBM Enterprise COBOL Version 4 Release 2
Performance Tuning".

I'm imagining that it is an "online" application. Can you estimate how many groups of calculations between ENTER and and returning the result in a variety of cases (as Phil has been encouraging)? Blast a couple of hundred thousand calcs in a little batch program, and see how much CPU you use. It'll give you some idea when you scale it down.

It may be a bigger task to convince everyone (including youself) that the calculations lead to the results expected by the user. As an example, for fun, do the same thing (carefully, lots of parentheses) as one big compute, and compare the answers (to the calculations) and the run times.

How is the user doing this currently? It is one thing to have the definition of the calculation, but you might have to get to how that definition is implemented currently to be able to get your answers the same.

Definitely follow Bill's advice about looking at the generated code. Look particularly at your WK-CALC05, I think you should seperate that out. Look at the intermediate fields the compiler generates.

I don't know what this

delago · New User Joined: 29 Jul 2005 Posts: 21 Location: Brazil

Hi Phrzby Phil.

The is very difficult to mensure exactly the quantity of calcs I executing on the day.
I have a estimate :
On-line : run 19500 SQL statements, where for each calculating I need to do 3 calcs to 1 result.
19500 x 3 = 58500 mathematical operations (estimated).

This operation is executed 20.000 for day.

Now, I alter the process to decrease a number of SQL statments for better performance, but I need some ideias for the future implementations note degrade the performance.

Tks

Fernando Delago - Brazil/São Paulo

Phrzby Phil · Posted: Wed Apr 20, 2011 8:36 pm

I just wanted to be sure that you have enough calcs per day to make computer efficiency more important than human readability.

Human meaning, of course, not just you, the author, but also the entry-level person who will inherit this at 3am when you are on vacation or have been promoted.