IBM has spent a lot of resources developing ILE RPG as a means of extending the venerable language into the 21st century. With ILE RPG, IBM says you can build modular RPG programs that use routines written and compiled in other languages on the AS/400. But most importantly, the implication from IBM is that the performance is very good. With this in mind, many shops are rigorously exploring ILE RPG in an effort to reengineer their applications for greater flexibility and performance.
Well, now that all the smoke has cleared from the glossy presentations about ILE, how good is it? Here are some tests I ran and my conclusions from looking at the numbers. The results were somewhat surprising.
I ran the tests shown in 1 on a dedicated E02 system using V3R1. IBMhas not verified the results in 2, but the RPG-to-RPG tests approximate the performance numbers shown in a COMMON presentation, "ILE Performance Comparisons."
I ran the tests shown in Figure 1 on a dedicated E02 system using V3R1. IBMhas not verified the results in Figure 2, but the RPG-to-RPG tests approximate the performance numbers shown in a COMMON presentation, "ILE Performance Comparisons."
The function I performed was moving one byte of information using a variety of methods. I used various iterations by submitting batch jobs. The numbers shown are in seconds and represent the CPU times for the jobs as reported by the system Job Accounting function.
Looking strictly at the numbers, it's not too difficult to draw the following conclusions:
o If you have 100 things to do, do it any way you want. ILE provides no significant advantage or disadvantage to the performance of the executing program.
o If you have 1,000 things to do, it's a big "so what" for all solutions except LR on and calling from RPG to CL. The other differences in performance are minor.
o If you have 1,000,000 things to do, even CALLB (test 4) will add some overhead. Assume you have a job that takes 20 minutes using EXSR (test 1). If CALLB adds 22 seconds for 100,000 iterations, then it would add 220 seconds for 1,000,000.
o If you read 1,000,000 records and do five CALLBs per record, the job that takes 20 minutes using EXSR will take roughly 40 minutes using CALLB. That's twice the performance hit for the price of a "modular" solution.
Looking at the results in more detail leads me to the following analysis:
Using the normal CALL command from RPG to CL is faster than using an ILE program and a CALLB to a CL module (compare jobs 5 and 6). This is mostly because CL has no concept like RPG's LR off, and, consequently, the overhead that occurs on any call to CL to set up the program still exists.
Because the performance of an ILE CALLPRC command from CL to RPG is very good, it may make some sense in RPG programs that need CL functions to have an ILE program with a mainline CL module that is returned to from RPG when a CL function is needed. In this case, this method may perform better than calling a CL program or a CALLB to a CL module.
I was impressed by how IBM lowered the overhead by having an RPG module calling another RPGmodule versus the traditional approach of calling an RPG program. Unfortunately, as my test results show, there is still a significant overhead of calling a module versus the traditional method of using an RPG subroutine.
There may be some more things that IBM can do to lower some of the overhead, but at this point, I would reach the same conclusion that was reached in the early days of the S/38: If you have big volume and you want to run as fast as possible, don't modularize. On the other hand, if you have a subfunction you only do 100 times a day, it doesn't matter how you do it.
Jim Sloan, president of Jim Sloan, Inc., is a software vendor and consultant. A retired IBMer, Sloan was a software planner on the S/38 when it began as a piece of paper. He also worked on the planning and early releases of the AS/400. In addition, Sloan wrote the TAA tools that exist in QUSRTOOL and is now the owner of the TAA Productivity Tools product. He has been a speaker at COMMON for many years.
Lab Notes: ILE Call Performance
Figure 1: Test Descriptions
Test Description 1. RPG program with subroutine An RPG program executed a subroutine. 2. RPG program to RPG program-LR on An RPG program used the CALL operation to call another RPG program that ended with LR on. 3. RPG program to RPG program-LR off An RPG program used the CALL operation to call another RPG program that ended with LR off. 4. RPG module to RPG module-LR off An RPG module used the CALLB operation to call another RPG module that ended with LR off. 5. RPG program to CL program An RPG program used the CALL operation to call a CL program. 6. RPG module to CL module An ILE program invoked an RPG module that used the CALLB operation to call a CL module. 7. CL program to RPG program-LR off A CL program used the CALL command to call an RPG program that ended with LR off. 8. CL module to RPG module-LR off An ILE program invoked a CL module that used the CALLPRC command to call an RPG module that ended with LR off.
Lab Notes: ILE Call Performance
Figure 2: Test Results in CPUSeconds
Test 100 1000 50,000 100,000 Iterations Iterations Iterations Iterations 1. RPG program with subroutine .775 .851 3.970 7.481 2. RPG program to RPG program-LR on 2.868 21.591 1040.191 2074.436 3. RPG program to RPG program-LR off .931 2.132 67.747 134.328 4. RPG module to RPG module-LR off 1.199 1.435 15.177 29.455 5. RPG program to CL program 1.576 8.666 395.142 789.490 6. RPG module to CL module 2.057 9.958 438.860 876.653 7. CL program to RPG program-LR off 1.961 15.397 749.208 1490.823 8. CL program to RPG module-LR off .820 1.170 19.703 38.625
LATEST COMMENTS
MC Press Online