Sparkling Water Tuning¶
For running Sparkling Water, general recommendations include:
Increase available memory in the driver and executors (options
spark.driver.memory
resp.,spark.yarn.am.memory
andspark.executor.memory
).Make cluster homogeneous. Use the same value for driver and executor memory.
Increase PermGen size if you are running on top of Java7 (options
spark.driver.extraJavaOptions
resp.,spark.yarn.am.extraJavaOptions
andspark.executor.extraJavaOptions
).In rare cases, it helps to increase
spark.yarn.driver.memoryOverhead
,spark.yarn.am.memoryOverhead
, orspark.yarn.executor.memoryOverhead
.
For running Sparkling Water on top of YARN:
Make sure that YARN provides stable containers; do not use preemptive YARN scheduler.
Make sure that the Spark application manager has enough memory, and increase PermGen size.
In the case of a container failure, YARN should not restart the container, and the application should gracefully terminate.
Furthermore, we recommend that you configure the following Spark properties to speed up and stabilize the creation of H2O services on top of Spark cluster:
Property |
Value |
Explanation |
---|---|---|
All environments (YARN/Standalone/Local) |
||
|
|
Number of seconds to wait for task launched on data-local node. We recommend to increase since we would like to make sure that H2O tasks are processed locally with data. |
|
|
Make sure that Spark starts scheduling when it sees 100% of resources. |
|
|
Do not try to retry failed tasks. |
|
|
Increase PermGem if you are running in Java7 on the Spark driver. |
|
|
Increase PermGem if you are running in Java7 on the Spark executor. |
|
|
Interval between each
executor heartbeats to
the driver. This property
should be significantly
less than
|
YARN environment |
||
|
|
Disable Spark support for dynamic allocation. |
|
|
Increase PermGem if you are running in Java7 on the Yarn application master. |
|
increase |
Increase memory overhead if it’s necessary of the container with driver node. |
|
increase |
Increase memory overhead if it’s necessary of the containers with executor nodes. |
|
increase |
Increase memory overhead if it’s necessary of the Yarn application master. |
|
|
Do not try to restart executors after failure and directly fail the computation. |