There are memory utilization issues with all versions of the Usage Meter prior to version 3.0.
We correctly most of these issues by having the collections run serial instead of parallel by default, modified the swap size and other configurations, autoconfigure JVM memory for appliance on power up based on memory allocated for virtual machine, expose JMX for monitoring memory utilization and other stats, and added health check tools to restart services if they appear to not be functioning properly.
Are you in a position where you cannot upgrade? If so, we can pull together some configurations to make on 2.3 that should help with your uptime.