[HADOOP-14960] Add GC time percentage monitor/alerter

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.0.0, 2.10.0
• Hadoop Flags: Reviewed
• Assignee: Misha Dmitriev
• Reporter: Misha Dmitriev
• Created: 18/Oct/17 18:57
• Updated: 20/Feb/20 23:37
• Resolved: 10/Nov/17 05:08

Description

Currently class org.apache.hadoop.metrics2.source.JvmMetrics provides several metrics related to GC. Unfortunately, all these metrics are not as useful as they could be, because they don't answer the first and most important question related to GC and JVM health: what percentage of time my JVM is paused in GC? This percentage, calculated as the sum of the GC pauses over some period, like 1 minute, divided by that period - is the most convenient measure of the GC health because:

• it is just one number, and it's clear that, say, 1..5% is good, but 80..90% is really bad
• it allows for easy apple-to-apple comparison between runs, even between different apps
• when this metric reaches some critical value like 70%, it almost always indicates a "GC death spiral", from which the app can recover only if it drops some task(s) etc.

The existing "total GC time", "total number of GCs" etc. metrics only give numbers that can be used to rougly estimate this percentage. Thus it is suggested to add a new metric to this class, and possibly allow users to register handlers that will be automatically invoked if this metric reaches the specified threshold.

Issue Links

• breaks: HADOOP-15067 GC time percentage reported in JvmMetrics should be a gauge, not counter
• is related to: HDFS-15176 Enable GcTimePercentage Metric in NameNode's JvmMetrics.
• is related to: HBASE-19238 Port Hadoop's new GcTimeMonitor