Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Fixed
-
0.17.1
-
None
-
None
Description
On a cluster with about 1700 nodes, when a job with about 100,000 maps and 10,000 reduces completed, the JobTracker, even with 80 handlers, could not handle the rpc call load during promotion of the job, such that at the end, because of the discarded heartbeats, the JobTracker lost nearly all TaskTrackers (about 10 TaskTrackers left). Promotion took more than 40 minutes.
They reconnected and everything recovered, but this might have been just luck.
Shouldn't there be an adaptive throttling of the rate in heartbeats and TaskCompletionEvents?
Sample messsages:
2008-07-22 18:21:55,831 WARN org.apache.hadoop.ipc.Server: Call queue overflow discarding oldest call heartbeat(org.apache.hadoop.mapred.TaskTrackerStatus@115f6b6, false, true, 18137) from xxx
2008-07-22 18:21:55,834WARN org.apache.hadoop.ipc.Server: Call queue overflow discarding oldest call getTaskCompletionEvents(job_200807190635_0012, 119567, 50) from yyy
...
2008-07-22 19:02:28,821 WARN org.apache.hadoop.ipc.Server: IPC Server handler 1 on 9020, call heartbeat(org.apache.hadoop.mapred.TaskTrackerStatus@19d32fa, false, true, 18199) from zzz: discarded for being too old (40936)
2008-07-22 19:02:28,821 WARN org.apache.hadoop.ipc.Server: IPC Server handler 34 on 9020, call getTaskCompletionEvents(job_200807190635_0012, 119567, 50) from uuu: discarded for being too old (40978)