Originally posted by: ankushg

Hi All,

I am looking to perfomance tune maps on our application. The platform is Z/OS - CICS. I want to start with map profiler performance tune the code. I have a couple of question:

1. Apart from any changing LOOKUP & EXTRACT type functions to SEARCHUP and SEARCHDOWN (when series is sorted) are there any other TIPS & TRICKS to improve performance by changing mapping rules (something like removing unwanted IF conditions).

2. What is the unit of time figures in the profiler(10000th of a second)? If yes, would changing the mapping rules provide any performance gains on CICS?

Thanks in Advance
#IBM-Websphere-Transformation-Extender
#IBMSterlingTransformationExtender
#DataExchange

Originally posted by: phil2030

I have to ask are your maps running that slowly? Presumably you have a perceived performance problem, is it identifying the input or producing the output? SEARCHUP/DOWN are good if your series is substantial, but are more restrictive in terms of search criteria and don't forget the overhead of maintaining the sort order. Type trees can often be simplified (just block out data you ar not interested in, don't bother with defining the entire complex structure), reorder partitioned objects so that more frequently occuring objects are nearer the top of the list. However, as usual it depends on what you want to achieve in terms of output as to what you can and can't do - only you can decide really.
I don't know about the resolution of the timer, I guess it might change per platform. That said, the short the time spent in any function and the fewer calls to each function you can get away with on any platform would presumably be the goal.
Personally, I might run the profiler if I thought a map was particularly slow (or just to get a final feel on performance prior to testing) and focus on any obvious slow areas, but wouldn't bother otherwise.
It goes without saying not to perform any tracing and try your workarea in memory or put it on the fastest drive available.
#DataExchange
#IBM-Websphere-Transformation-Extender
#IBMSterlingTransformationExtender

Originally posted by: SystemAdmin

Another thing to check is perhaps memory allocation for the user running the map. I vaguely recall an issue where a map that took minutes on the desktop was timing out when run on the mainframe. Don't recall if it was CICS or DOS or what exactly the solution was. I suspect Chris Hull, IBM WTX Second Level Support, may was the person who assisted us on resolving our performance issue.
#DataExchange
#IBMSterlingTransformationExtender
#IBM-Websphere-Transformation-Extender

Originally posted by: ankushg

Thanks for the replies.

Phil, there is no specific performance problem I am looking to fix during this exercise. Just looking to improve the performance generally (if not already optimized). "Having the more likely objects occur first in the partition" are the kind of ideas I am looking for. However, we cant change typetrees for standard formats eg SWIFT or any other payment messages.

The memory allocation is default (Workspace: file, pagesize 64 and pagecount 8). The instruction we pass to the map during a single call is fairly small.(avg size 4k). I suppose we wouldnt need to change that.

Recently, after running the profiler and analysing the maps, we removed a few unneccesary IF conditions. Following were the observations:
1. The profiler produces different figures for the same data on the same map when run at different times. First run showed 400 units and the second run showed 500 units. Why should this happen???

2. After removing the IF conditions, a couple of times the profiler should an improvement of 100 units. Does this 100 units means anything keeping in mind that the profiler timings vary in different runs??
#IBMSterlingTransformationExtender
#DataExchange
#IBM-Websphere-Transformation-Extender

Originally posted by: phil2030

If your getting different runtimes with the same data you want to run the map 'n' times ('n' being a suitably large number of your choice) and take the average. The variation is probably down to the machine busy doing "other things" at the time.
If your talking 1000ths of a second then you are suggesting a .1s difference. To me I would class it as fairly irrelevant and wouldn't bother - depends on how you're threading it, message volumes, any SLA and what else the machine is tasked for. I am guessing you are running this from the command server on this platform?
In answer to your #2 query, is that consistently 100 quicker or just on the one time you ran it. It sounds like there is a "margin of error" here and with 4K source sizes, your not going to be handling the sizes required to show any significant performance advantages when "tweaking" rules.
Unless it is obviously slow compared to what other testing you have done, I'd leave it...
#DataExchange
#IBM-Websphere-Transformation-Extender
#IBMSterlingTransformationExtender