Improve memory usage for IRAC

Asynchronous cross-site replication (codename IRAC) uses a vector clock internally to detect conflicts, and this vector clock is just a simple map between the site name and the version.

The "possible" issue is when the primary owner backups the data to the remote site, a new String instance is created for each vector clock position. While the site's name may be small, it may take a large portion of the heap space when you store millions of entries in memory.

This situation can be improved by caching the site names (usually, only 2 or 3 sites are used 99% of the time).