What enabled the increase in the replication service throughput from about 300 to 700 messages per second? How were ConcurrentDictionary and SemaphoreSlim used?