All

Congratulations to Werner Rams for the first solution to problem #17.  His solution compresses and decompresses 3 records using CLCL to detect end of duplicate character strings.  This solution executed 30,629 instructions.  To view the source code and the generated log file running on z390 visit:

http://z390.sourceforge.net/z390_Mainframe_Assemble_Coding_Contest.htm

I’ve added a new problem # 18 to measure the performance gain from using new IBM z10 opcode CIJNE:

Write a benchmark program to calculate the percent  performance improvement to 2 decimal places when replacing the following loop code:

LOOP DS 0H
            BCTR      R1,0
*** APPLICATION CODE COMMENTED OUT FOR TEST ***
            LTR         R1,R1
            JNE         LOOP

with the following optimized loop code using the new z10 compare immediate and branch relative opcode code:

LOOP DS 0H
            BCTR    R1,0
*** APPLICATION CODE COMMENTED OUT FOR TEST ***
            CIJNE  R1,0,LOOP

The performance improvement in this case comes from replacing 2 instruction cycles fetching a total of 6 bytes with a single instruction cycle fetching 6 bytes.  You can use whatever interval timing method is available on your system such as TIME BIN (requires running standalone).  The initial values in R1 must be set to perform enough iterations to reduce the timing error due to interval timer precision etc.  To code and unit test solution on z390 you will need version v1.4.01 with the new z10 opcodes scheduled for release by 03/14/08.  To run the real test, you will need an IBM z10 mainframe and updated HLASM.  My own initial test on pre-release version of z390 v1.4.01 indicates about a 15% improvement but this is measuring J2SE VM emulation overhead of each instruction cycle on Intel Dual-Core chip versus measuring the IBM z10 hardware/Millicode.  I'll be very interested to hear the real results.

Don Higgins

mailto:don@...

http://don.higgins.net