I solved this by coercing
(get-internal-real-time)
into a double float.

Looks like the epoch start time for these functions is different from lisp to lisp.

ACL: (get-internal-real-time) => 8456209511
ABCL: (get-internal-real-time) => 1401814359707

Does anyone know the date from which the value for (get-internal-real-time) is computed for abcl?
bonasso <bonasso@traclabs.com> writes:

The RAPs system I'm running in abcl makes extensive use of the
difference between two sequential get-internal-real-time calls. But I
get significantly different results in abcl than in acl.

Here's an example:

In acl:

CL-USER(24): (setf foo (/ (get-internal-real-time)(float
internal-time-units-per-second)))
8212185.5
CL-USER(25): (setf bar (/ (get-internal-real-time)(float
internal-time-units-per-second)))
8212195.0
CL-USER(26): (- bar foo)
9.5 ;; nine+ seconds

In abcl:
CL-USER(15): (setf foo (/ (get-internal-real-time)(float
internal-time-units-per-second)))
1.40157018E9
CL-USER(16): (setf bar (/ (get-internal-real-time)(float
internal-time-units-per-second)))
1.40157018E9
CL-USER(17): (- bar foo)
0.0 ;; 0 seconds

Thus, lots of time has to elapse between the calls before any
difference is produced. (/ (get-internal-real-time)(float
internal-time-units-per-second)) is the primitive time call used in
the RAPs system.

Of course the first thing I noticed is that the single floats in acl
don't have the E notation.
Is there a setting in abcl I can use to get the same behavior as in
acl, or is there another problem going on?

(get-internal-real-time) => 1401731099641, that's too big to be
represented precisely as a single float. double float will work better.
And better divide by internal-time-units-per-second before doing
division. Even better would be to avoid converting into floats.