Apache Hadoop Yarn - Underutilization of cores

Question

No matter how much I tinker with the settings in yarn-site.xml i.e using all of the below options

yarn.scheduler.minimum-allocation-vcores
yarn.nodemanager.resource.memory-mb
yarn.nodemanager.resource.cpu-vcores
yarn.scheduler.maximum-allocation-mb
yarn.scheduler.maximum-allocation-vcores

i just still cannot get my application i.e Spark to utilize all the cores on the cluster. The spark executors seem to be correctly taking up all the available memory, but each executor just keeps taking a single core and no more.

Here are the options configured in spark-defaults.conf

spark.executor.cores                    3
spark.executor.memory                   5100m
spark.yarn.executor.memoryOverhead      800
spark.driver.memory                     2g
spark.yarn.driver.memoryOverhead        400
spark.executor.instances                28
spark.reducer.maxMbInFlight             120
spark.shuffle.file.buffer.kb            200

Notice that spark.executor.cores is set to 3, but it doesn't work. How do i fix this?

Winwaloe

Answer 1

The problem lies not with yarn-site.xml or spark-defaults.conf but actually with the resource calculator that assigns the cores to the executors or in the case of MapReduce jobs, to the Mappers/Reducers.

The default resource calculator i.e org.apache.hadoop.yarn.util.resource.DefaultResourceCalculator uses only memory information for allocating containers and CPU scheduling is not enabled by default. To use both memory as well as the CPU, the resource calculator needs to be changed to org.apache.hadoop.yarn.util.resource.DominantResourceCalculator in the capacity-scheduler.xml file.

Here's what needs to change.

<property>
    <name>yarn.scheduler.capacity.resource-calculator</name>
    <value>org.apache.hadoop.yarn.util.resource.DominantResourceCalculator</value>
</property>

Winwaloe

well, stackoverflow encourages you to share one's knowledge, so doing just that :) – Winwaloe

good work - I just went through a whole lot of trouble finding out exactly the same, including filing a bug with yarn (it is not a bug). Just wish I found this earlier. See official Hadoop documentation. Careful when copy-pasting, I also found out that the original documentation misspelt Resource as Resourse within the configuration (s vs c). – Griddlecake

for me over utilisation of cpu is coming for mapreduce job... only 2 cores configured,but load avarage for mapreduce is high.. some time a hike in cpu % is also there.. any idea? – Twenty

But does it really uses only one core per executor? Or it is just the calculation for the number of executors? I mean, does --executor-cores value still work when using the DefaultResourceCalculator? – Prowl

Answer 2

I had the similar kind of issue and from my code i was setting up the spark.executor.cores as 5. Even though it was just taking 1 which is the default core. In the spark UI and in environment tab i was seeing 5 cores. But while checking the executors tabs i was just able to see 1 process is in RUNNING state against an executor. I was using the spark version 1.6.3.

So then i have tried to hit the spark-submit command as --conf spark.executor.cores=5 which is working fine as using 5 cores

or just

--executor-cores 5 which also works.

Minier