Hi, Apologies if this is a repeat. I don't know if my earlier posting got out due to the timing of my subscription being accepted.
I'm a new subscriber to this list with some questions about Hadoop JMX metrics. We are running version 0.20.2. Hadoop-metrics.properties is configured for GangliaContext31 for dfs, mapred, and rpc, with period=10. Hadoop-env.sh is setup for jmxremote, with namenode and datanode accessible on ports 8004 & 8005. I am able to access the MBeans and retrieve attributes and values - using both jconsole and my own java client code. These are my questions/comments: 1. (Comment) It seems that certain counters such as number of reads and writes are interval values - they reset to zero after publication, in my case, every 10 seconds. For a JMX client querying these stats at arbitrary, or even regular, time intevals, this would not be a good method for deriving useful info from these stats. Would NullContextWithUpdateThread context behave differently, that is, not reset? For a "dashboard"-type application, metrics would be meaningful only when receiving the periodic publications. 2. Our Hadoop runs on a 4-node cluster with 4 datanodes. The NameNodeInfo mbean (class org.apache.hadoop.hdfs.server.namenode.FSNamesystem) reports "used space" and "last contact" on all 4 nodes. How to interpret the value for "last contact"? For example, for a DeadNode, the last contact value is a large number. 3. There's a datanode mbean report with a key property that begins with "FSDatasetState-DS-...", (class org.apache.hadoop.hdfs.server.datanode.FSDataset) that contains counts for remaining/capacity/used space - but it's only for one node. Why is there not one for each datanode? 4. Is the spec for the GangliaContext publication format available, or is it proprietary? Maybe my issues are due to the version we're running? I'm sure things will be different with metrics2. Thanks for any information. David Wong HP
