Hi, guys, Recently we're looking into HDFS NameNode FSNamesystemLock implementation. As it uses Java's ReentrantReadWriteLock, after configing the lock to nonfair mode, the RPC throughput has been improved a lot. Just wonder whether anyboby have played around the reader/writer preference further in the lock (or directly used another lock implementation) to achieve higher throughput, absolutely there will be trade-off between throughput and write starvation.
thanks, Wei
