Issues when scaling XDCR on single node

Perry asked me recently about peak XDCR throughput for single node setup.

It made me curious enough to run following test using our the most fast servers:
1. Take single server with single default bucket
2. Load 5M items, wait for persistence
3. Enable XDC replication stream to another single server (so that 1-to-1 setup within the same LAN)
4. Wait for initial replication and monitor overall cluster performance.

As results I observer average throughput about 15K ops/sec. Not bad, absolutely.

However, both cluster don't look saturated at all (I mean h/w resources). Obviously I will try tuning options or reduced number of vbuckets. But basically where do we expect bottleneck in such setups?