Spark on Kubernetes: Kubernetes killing executors because of overallocation of memory

[Jayesh Lalwani <ja...@capitalone.com>]

We are running Spark 2.3 on a Kubernetes cluster. We have set the following
spark configuration options

"spark.executor.memory": "7g",
    "spark.driver.memory": "2g",
    "spark.memory.fraction": "0.75"

WHat we see is
a) In the SPark UI, 5G has been allocated to each executor, which makes
sense because we set spark.memory.fraction=0.75
b) Kubernetes reports the pod memory usage as 7.6G

WHen we run a lot of jobs on the Kubernetes cluster, Kubernetes starts
killing the executor pods, because it thinks that the pod is misbehaving.

We logged into a running pod, and ran the top command, and most of the 7.6G
is being allocated to the executor's java process

Why is Spark taking 7.6G instead of 7 G? Where is the 600MB being allocated
to? Is there some configuration that controls how much of the executor
memory gets allocated to Permgen vs the memory that gets allocated to the
heap?
________________________________________________________

The information contained in this e-mail is confidential and/or proprietary to Capital One and/or its affiliates and may only be used solely in performance of work or services for Capital One. The information transmitted herewith is intended only for use by the individual or entity to which it is addressed. If the reader of this message is not the intended recipient, you are hereby notified that any review, retransmission, dissemination, distribution, copying or other use of, or taking of any action in reliance upon this information is strictly prohibited. If you have received this communication in error, please contact the sender and delete the material from your computer.

Re: Spark on Kubernetes: Kubernetes killing executors because of overallocation of memory

[Matt Cheah <mc...@palantir.com>]

Hi there,

 

You may want to look at setting the memory overhead settings higher. Spark will then start containers with a higher memory limit (spark.executor.memory + spark.executor.memoryOverhead, to be exact) while the heap is still locked to spark.executor.memory. There’s some memory used by offheap storage from Spark that won’t be accounted for in just the heap size.

 

Hope this helps,

 

-Matt Cheah

 

From: Jayesh Lalwani <ja...@capitalone.com>
Date: Thursday, August 2, 2018 at 12:35 PM
To: "user@spark.apache.org" <us...@spark.apache.org>
Subject: Spark on Kubernetes: Kubernetes killing executors because of overallocation of memory

 

We are running Spark 2.3 on a Kubernetes cluster. We have set the following spark configuration options

"spark.executor.memory": "7g",

    "spark.driver.memory": "2g",

    "spark.memory.fraction": "0.75"

 

WHat we see is

a) In the SPark UI, 5G has been allocated to each executor, which makes sense because we set spark.memory.fraction=0.75
b) Kubernetes reports the pod memory usage as 7.6G

 

WHen we run a lot of jobs on the Kubernetes cluster, Kubernetes starts killing the executor pods, because it thinks that the pod is misbehaving.

 

We logged into a running pod, and ran the top command, and most of the 7.6G is being allocated to the executor's java process

 

Why is Spark taking 7.6G instead of 7 G? Where is the 600MB being allocated to? Is there some configuration that controls how much of the executor memory gets allocated to Permgen vs the memory that gets allocated to the heap?

 

 

 

The information contained in this e-mail is confidential and/or proprietary to Capital One and/or its affiliates and may only be used solely in performance of work or services for Capital One. The information transmitted herewith is intended only for use by the individual or entity to which it is addressed. If the reader of this message is not the intended recipient, you are hereby notified that any review, retransmission, dissemination, distribution, copying or other use of, or taking of any action in reliance upon this information is strictly prohibited. If you have received this communication in error, please contact the sender and delete the material from your computer.