Hi -
I have a three node Hadoop 3.1.0 cluster on which the daemons are
distributed like so:
Daemons on msba02a...
20112 NameNode
20240 DataNode
24101 JobHistoryServer
20918 WebAppProxyServer
20743 NodeManager
20476 SecondaryNameNode
Daemons on msba02b...
22547 DataNode
22734 ResourceManager
23007 NodeManager
Daemons on msba02c...
10005 NodeManager
9818 DataNode
All three nodes run Gentoo Linux and have either one or two volumes
devoted to HDFS; HDFS reports a size of 5.7TiB.
Previously, HDFS and MapReduce (testing with the archetypical
"wordcount" job on a 5.8GiB XML file) worked fine in an environment
where all three machines are on the same office LAN and get their IP
addresses from DHCP; dynamic DNS creates network host names based on the
machines' host names as reported by the machines' DHCP clients. FQDNs
were used for all intra- and inter-machine references in the Hadoop
configuration files.
Since then, I've changed things so that msba02a now has a second NIC
that connects to an independent LAN along with the other two machines
using their built-in NICs like before; msba02b and msba02c reach the
Internet by going through NAT on msba02a. /etc/hosts on all three
machines has been populated with the static IPs I gave them like so:
127.0.0.1 localhost
1.0.0.1 msba02a
1.0.0.10 msba02b
1.0.0.20 msba02c
So now if I shell into msba02a and run the wordcount job with the test
XML file sitting in HDFS with replication set to 3, the job *does* run
and gives me the expected output file...but the workload doesn't
distribute to all cores on all nodes like before; it all executes on
msba02a. In fact, it doesn't even run on all cores on msba02a; it seems
to light up just one core at any given moment. The job used to run on
the cluster in 1m48s; now it takes 5m56 (a ratio I can't understand;
these are all four-core, eight-thread machines so I'd expect a ratio of
close to 24:1, not 3:1). The only time the other two nodes light up at
all is near the end of the job when the output file (770MiB) is written
out to HDFS.
I've gone through
https://hadoop.apache.org/docs/current3/hadoop-project-dist/hadoop-hdfs/HdfsMultihoming.html
and set the values shown there to 1.0.0.1 in hdfs-site.xml on msba02a in
hopes of getting the daemons to bind to the cluster-facing NIC instead
of the outward-facing NIC, but it seems to me like HDFS is working
exactly like it's supposed to. Note that the ResourceManager daemon runs
on msba02b and therefore doesn't need to be bound to a particular NIC;
it still uses that machine's only NIC like before except now its IP
address is static and is resolved via its local /etc/hosts.
The only errors showing up in the daemon logs of any nodes seem to be
e.g.
"org.apache.hadoop.security.token.delegation.AbstractDelegationTokenSecretManager:
ExpiredTokenRemover received java.lang.InterruptedException: sleep
interrupted" in hadoop-yarn-resourcemanager-msba02b.log and
hadoop-mapred-historyserver-msba02a.log.
As for the hadoop run output, previously when everything was working
things would get to point where it would print out a series of lines like
map 0% reduce 0%
and that line would repeat with "map" percentage climbing first and then
the "reduce" percentage would climb until both numbers reached 100% and
the job would wrap up soon afterward. Now, it intersperses those lines
with other output and it skips around, like this:
*2018-06-05 17:45:34,338 INFO mapreduce.Job: map 100% reduce 0%*
2018-06-05 17:45:36,295 INFO mapred.MapTask: Finished spill 0
2018-06-05 17:45:36,295 INFO mapred.MapTask: (RESET) equator
61480136 kv 15370028(61480112) kvi 13480948(53923792)
2018-06-05 17:45:36,882 INFO mapred.MapTask: Spilling map output
2018-06-05 17:45:36,882 INFO mapred.MapTask: bufstart = 61480136;
bufend = 10372007; bufvoid = 104857566
2018-06-05 17:45:36,882 INFO mapred.MapTask: kvstart =
15370028(61480112); kvend = 7835876(31343504); length = 7534153/6553600
2018-06-05 17:45:36,882 INFO mapred.MapTask: (EQUATOR) 17997991 kvi
4499492(17997968)
2018-06-05 17:45:38,774 INFO mapred.MapTask: Finished spill 1
2018-06-05 17:45:38,774 INFO mapred.MapTask: (RESET) equator
17997991 kv 4499492(17997968) kvi 2642780(10571120)
2018-06-05 17:45:38,910 INFO mapred.LocalJobRunner:
2018-06-05 17:45:38,910 INFO mapred.MapTask: Starting flush of map
output
2018-06-05 17:45:38,910 INFO mapred.MapTask: Spilling map output
2018-06-05 17:45:38,911 INFO mapred.MapTask: bufstart = 17997991;
bufend = 40956853; bufvoid = 104857600
2018-06-05 17:45:38,911 INFO mapred.MapTask: kvstart =
4499492(17997968); kvend = 1327036(5308144); length = 3172457/6553600
*2018-06-05 17:45:39,340 INFO mapreduce.Job: map 4% reduce 0%*
2018-06-05 17:45:39,684 INFO mapred.MapTask: Finished spill 2
2018-06-05 17:45:39,788 INFO mapred.Merger: Merging 3 sorted segments
2018-06-05 17:45:39,788 INFO mapred.Merger: Down to the last
merge-pass, with 3 segments left of total size: 34645401 bytes
2018-06-05 17:45:40,251 INFO mapred.Task:
Task:attempt_local1155504279_0001_m_000002_0 is done. And is in the
process of committing
2018-06-05 17:45:40,253 INFO mapred.LocalJobRunner: map > sort
2018-06-05 17:45:40,253 INFO mapred.Task: Task
'attempt_local1155504279_0001_m_000002_0' done.
2018-06-05 17:45:40,253 INFO mapred.Task: Final Counters for
attempt_local1155504279_0001_m_000002_0: Counters: 23
File System Counters
FILE: Number of bytes read=106419805
FILE: Number of bytes written=202253153
FILE: Number of read operations=0
FILE: Number of large read operations=0
FILE: Number of write operations=0
HDFS: Number of bytes read=410006948
HDFS: Number of bytes written=0
HDFS: Number of read operations=9
HDFS: Number of large read operations=0
HDFS: Number of write operations=1
Map-Reduce Framework
Map input records=2653033
Map output records=4553651
Map output bytes=130562451
Map output materialized bytes=31060160
Input split bytes=95
Combine input records=5425504
Combine output records=1618222
Spilled Records=1618222
Failed Shuffles=0
Merged Map outputs=0
GC time elapsed (ms)=114
Total committed heap usage (bytes)=1301807104
File Input Format Counters
Bytes Read=134348800
2018-06-05 17:45:40,253 INFO mapred.LocalJobRunner: Finishing task:
attempt_local1155504279_0001_m_000002_0
2018-06-05 17:45:40,253 INFO mapred.LocalJobRunner: Starting task:
attempt_local1155504279_0001_m_000003_0
2018-06-05 17:45:40,254 INFO output.FileOutputCommitter: File Output
Committer Algorithm version is 2
2018-06-05 17:45:40,254 INFO output.FileOutputCommitter:
FileOutputCommitter skip cleanup _temporary folders under output
directory:false, ignore cleanup failures: false
2018-06-05 17:45:40,254 INFO mapred.Task: Using
ResourceCalculatorProcessTree : [ ]
2018-06-05 17:45:40,255 INFO mapred.MapTask: Processing split:
hdfs://msba02a:9000/allcat.xml:268435456+134217728
2018-06-05 17:45:40,265 INFO mapred.MapTask: (EQUATOR) 0 kvi
26214396(104857584)
2018-06-05 17:45:40,266 INFO mapred.MapTask:
mapreduce.task.io.sort.mb: 100
2018-06-05 17:45:40,266 INFO mapred.MapTask: soft limit at 83886080
2018-06-05 17:45:40,266 INFO mapred.MapTask: bufstart = 0; bufvoid =
104857600
2018-06-05 17:45:40,266 INFO mapred.MapTask: kvstart = 26214396;
length = 6553600
2018-06-05 17:45:40,266 INFO mapred.MapTask: Map output collector
class = org.apache.hadoop.mapred.MapTask$MapOutputBuffer
*2018-06-05 17:45:40,341 INFO mapreduce.Job: map 100% reduce 0%*
2018-06-05 17:45:41,079 INFO mapred.MapTask: Spilling map output
2018-06-05 17:45:41,079 INFO mapred.MapTask: bufstart = 0; bufend =
53799451; bufvoid = 104857600
2018-06-05 17:45:41,079 INFO mapred.MapTask: kvstart =
26214396(104857584); kvend = 18692744(74770976); length =
7521653/6553600
2018-06-05 17:45:41,079 INFO mapred.MapTask: (EQUATOR) 61425451 kvi
15356356(61425424)
2018-06-05 17:45:43,110 INFO mapred.MapTask: Finished spill 0
2018-06-05 17:45:43,110 INFO mapred.MapTask: (RESET) equator
61425451 kv 15356356(61425424) kvi 13514352(54057408)
2018-06-05 17:45:43,687 INFO mapred.MapTask: Spilling map output
2018-06-05 17:45:43,687 INFO mapred.MapTask: bufstart = 61425451;
bufend = 10294846; bufvoid = 104857586
2018-06-05 17:45:43,687 INFO mapred.MapTask: kvstart =
15356356(61425424); kvend = 7816592(31266368); length = 7539765/6553600
2018-06-05 17:45:43,687 INFO mapred.MapTask: (EQUATOR) 17920846 kvi
4480204(17920816)
2018-06-05 17:45:46,275 INFO mapred.MapTask: Finished spill 1
2018-06-05 17:45:46,275 INFO mapred.MapTask: (RESET) equator
17920846 kv 4480204(17920816) kvi 2573716(10294864)
2018-06-05 17:45:46,423 INFO mapred.LocalJobRunner:
2018-06-05 17:45:46,423 INFO mapred.MapTask: Starting flush of map
output
2018-06-05 17:45:46,423 INFO mapred.MapTask: Spilling map output
2018-06-05 17:45:46,423 INFO mapred.MapTask: bufstart = 17920846;
bufend = 41420321; bufvoid = 104857600
2018-06-05 17:45:46,423 INFO mapred.MapTask: kvstart =
4480204(17920816); kvend = 1126824(4507296); length = 3353381/6553600
Any hints as to why work isn't distributing? It seems to me like this
kind of network configuration for Hadoop clusters would be more the norm
than one where all nodes are on a network with everything else in an
environment (in our situation one driver for having cluster traffic
isolated is because the data files used may contain NDA-bound data that
shouldn't travel the office LAN unencrypted).
Thanks!