David Rorke created IMPALA-15187:
------------------------------------
Summary: TestIcebergTrinoInterop trino container startup failing
from excessive mem allocation request
Key: IMPALA-15187
URL: https://issues.apache.org/jira/browse/IMPALA-15187
Project: IMPALA
Issue Type: Bug
Components: Infrastructure
Reporter: David Rorke
Assignee: David Rorke
The Trino container is failing to start during TestIcebergTrinoInterop because
the JVM is attempting to commit 50GB of memory during startup:
{noformat}
OpenJDK 64-Bit Server VM warning: INFO: os::commit_memory(0x0000fff2f6000000,
53552873472, 0) failed; error='Not enough space' (errno=12) # # There is
insufficient memory for the Java Runtime Environment to continue. # Native
memory allocation (mmap) failed to map 53552873472 bytes. Error detail:
committing reserved memory.{noformat}
The reason appears to be the following in the jvm.config file for the Docker
image which tells the JVM to allocate 80% of the total container memory at
startup:
{noformat}
-XX:InitialRAMPercentage=80
-XX:MaxRAMPercentage=80{noformat}
There are no container memory limits so these percentages will be applied to
the total host memory (64GB in this case). minicluster_trino should have a
much lower memory requirement than this and these percentages could be reduced
to something like 10% initial and 30% max.
So far we've only seen this failure on ARM (not on Intel). The reason isn't
clear (some arm instances use a large 64KB page size so maybe more
fragmentation), but in any case the 50GB allocation is excessive and should be
reduced on both ARM and Intel.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
