I have been running a 2000 node cluster and measuring namenode performance. There are quite a few "Calls dropped" messages in the namenode log. The namenode machine has 4 CPUs and each CPU is about 30% busy. Profiling the namenode shows that the methods the consume CPU the most are addStoredBlock() and getAdditionalBlock(). The first method in invoked when a datanode confirms the presence of a newly created block. The second method in invoked when a DFSClient request a new block for a file.
I am attaching two files that were generated by the profiler. serverThreads40.html captures the scenario when the namenode had 40 server handler threads. serverThreads1.html is with 1 server handler thread (with a max_queue_size of 4000).
In the case when there are 40 handler threads, the total elapsed time taken by FSNamesystem.getAdditionalBlock() is 1957 seconds whereas the methods that that it invokes (chooseTarget) takes only about 97 seconds. FSNamesystem.getAdditionalBlock is blocked on the global FSNamesystem lock for all those 1860 seconds.
My proposal is to implement a finer grain locking model in the namenode. The FSNamesystem has a few important data structures, e.g. blocksMap, datanodeMap, leases, neededReplication, pendingCreates, heartbeats, etc. Many of these data structures already have their own lock. My proposal is to have a lock for each one of these data structures. The individual lock will protect the integrity of the contents of the data structure that it protects. The global FSNamesystem lock is still needed to maintain consistency across different data structures.
If we implement the above proposal, both addStoredBlock() and getAdditionalBlock() does not need to hold the global FSNamesystem lock. startFile() and closeFile() still needs to acquire the global FSNamesystem lock because it needs to ensure consistency across multiple data structures.