JobTracker running out of heap space

The JobTracker on our 0.17.0 installation runs out of heap space rather quickly, with less than 100 jobs (at one time even after just 16 jobs).
Running in 64-bit mode with larger heap space does not help – it will use up all available RAM.

2008-06-28 05:17:06,661 INFO org.apache.hadoop.ipc.Server: IPC Server handler 62 on 9020, call he
artbeat(org.apache.hadoop.mapred.TaskTrackerStatus@6f81c6, false, true, 17384) from xxx.xxx.xxx.xxx
:51802: error: java.io.IOException: java.lang.OutOfMemoryError: GC overhead limit exceeded
java.io.IOException: java.lang.OutOfMemoryError: GC overhead limit exceeded