22
As other reply suggests, COMP means big endian binary. COMP-3 is packed decimal- which means one decimal digit is mapped to each nibble.
I am not sure the previous reply got the issue around precision correct though.
PIC S9(9)V9(9) COMP and PIC S9(9)V9(9) COMP-3
Have exactly the same precision. That is part of the ANSI85 standard. It is the job of the compiler and runtime to ensure that the binary representation in the COMP has the appropriate transformations placed upon it to ensure exactly the same results are achieved as would be if usage was display or COMP-3.
IBM mainframe computers have packed decimal calculations in hardware. This is very helpful, because the conversion of decimal to binary scales as n squared n is the length of the number. This means that COMP-3 is every often the fastest format of the mainframe, but is less likely to be on distributed systems. However, this again is not always the case. For example, the Micro Focus native COBOL solution will tend to be faster in COMP-3 than COMP-5 for very large decimal precision (>18 digits) but the reverse for otherwise. The Managed COBOL system from Micro Focus is almost always fastest in COMP (actually, COMP-5 is the best - which is similar to COMP but will have hardware endian rather than enforcing big-endian memory layout).
Finally, my I suggest that for intermediate values and general mathematics, the newer data definitions of binary-long and binary-double are a better choice because then the compiler can make the decisions about how to store and optimize for you.
For more on COBOL on distributed and Managed COBOL check out this knol: http://knol.google.com/k/alex-turner/micro-focus-managed-cobol/2246polgkyjfl/4 and also feel free to look up cobol on facebook :)