The IBM JVM dump mechanism implemented by the dump agents uses a sophisticated strategy to determine where to save / store the dump files. The order of paths which are checked when the JVM tries to write the dumps is documented in the JVM Knowledge Center.
While this mechanism makes perfectly sense to ensure that the required debug data are written there is one drawback namely that it might fill either WebSphere's working directory (which is usually the <profile_home>) or the operating system's temp directory (/tmp by default on Unix systems). If any of those directories become full this might have a negative impact on Websphere itself of even on the whole running machine.
To avoid that the JVM uses the implemented fall-back mechanism and therefore to avoid that the operating system's temp directory gets filled up with JVM dumps you can use the following dump option: -Xdump:nofailover.