[
https://issues.apache.org/jira/browse/SPARK-54449?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Yifeng Wang updated SPARK-54449:
--------------------------------
Attachment: image-2025-11-21-22-03-25-116.png
> Spark UI display unexpected "Storage Memory" capacity when
> spark.memory.offHeap.enabled is false
> ------------------------------------------------------------------------------------------------
>
> Key: SPARK-54449
> URL: https://issues.apache.org/jira/browse/SPARK-54449
> Project: Spark
> Issue Type: Bug
> Components: Block Manager
> Affects Versions: 3.5.2
> Environment: Tested with Spark v3.5.2. Steps to reproduce included in
> below Description section.
>
>
> Reporter: Yifeng Wang
> Priority: Critical
> Attachments: image-2025-11-21-21-49-17-534.png,
> image-2025-11-21-21-49-34-881.png, image-2025-11-21-21-51-10-957.png,
> image-2025-11-21-21-51-39-197.png, image-2025-11-21-21-53-37-637.png,
> image-2025-11-21-21-53-59-322.png, image-2025-11-21-21-54-03-317.png,
> image-2025-11-21-22-00-03-629.png, image-2025-11-21-22-03-25-116.png,
> image-2025-11-21-22-03-34-768.png
>
>
> Dear Spark Community:
> *Current Behavior:* In the Spark UI (Executors tab) and Spark History Server,
> the "Storage Memory" column displays the total capacity as the sum of On-Heap
> Storage Memory + Off-Heap Storage Memory. Even when the configuration
> {{spark.memory.offHeap.enabled}} is explicitly set to {{{}false{}}}, the UI
> still adds the value of {{spark.memory.offHeap.size}} to the total displayed
> capacity.
> *Expected Behavior:* Perhaps when {{spark.memory.offHeap.enabled}} is set to
> {{{}false{}}}, the "Storage Memory" total in the UI should *only* reflect the
> On-Heap Storage Memory? The {{spark.memory.offHeap.size}} configuration
> should be ignored in the UI display calculation, similar to how it is ignored
> in YARN resource allocation logic.
>
> *Personal Understandings:* The issue stems from how {{BlockManager}} reports
> memory to the {{{}BlockManagerMaster{}}}. {{UnifiedMemoryManager}}
> initializes {{maxOffHeapMemory}} based on the configuration
> {{spark.memory.offHeap.size}} regardless of the {{enabled}} flag. Then
> {{BlockManager}} reads this value via
> {{memoryManager.maxOffHeapStorageMemory}} and passes it to the
> {{registerBlockManager}} RPC call. {{At register}} method, the total memory
> is calculated by simply adding on-heap and off-heap values without checking
> if off-heap is enabled.
> At ther parts of the codebase, such as {{Client.scala}} (YARN) and
> {{{}ResourceProfile.scala{}}}, which correctly utilize
> {{Utils.checkOffHeapEnabled}} to ensure off-heap memory is treated as 0 when
> the feature is disabled.
> *Proposed Fix:* Perhaps we need enforce the {{checkOffHeapEnabled}} logic
> before aggregating the total memory for the UI?
> !image-2025-11-21-21-49-34-881.png!
>
> *Steps to Reproduce:*
> # Configure a Spark application with the following settings:
> *
> ** {{spark.memory.offHeap.enabled=false}}
> *
> ** {{spark.memory.offHeap.size=10g}} (or any non-zero value)
>
> *Screenshots:*
> 1. Comment out configs
>
> Got: (Storage Memory =2 GiB)
>
>
>
> 2. Enable configs
>
> Got: (Storage Memory =12 GiB)
>
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
