Solution: Floating-point assist faults error
Most of these errors can be eliminated by compiling the offending application with the following flags:

-O2 -ftz

The -ftz (flush to zero) is an Intel specific flag and is probably not present in gcc.

However, since the floating-point assist faults do have an effect on performance, the customer should recompile the application with a '-g' to get the debugging information built in, then use gdb and the instruction pointer provided by the error message to find the section of source code that is producing the underflow (or the divide by zero or whatever is causing the floating-point assist fault.)

NOTE: The number of faults in /var/log/messages may not correlate with the number of faults being generated by an application as only 5 of these errors will be printed per second. For information on silencing these messages, see the section entitled "Floating Point Assist Warnings from Applications" in Chapter 1 of the Linux Configuration and Operations guide.

Problem Description
Floating-point assist faults error
Error: "kernel: <app name>(<PID>): floating-point assist fault at ip <address>, isr <address>"
Errors in /var/log/messages: "floating-point assist fault at ip 2000000000075bb2, isr 0000020000007001"