LuciferYang opened a new issue, #12616:
URL: https://github.com/apache/gluten/issues/12616
### Backend
VL (Velox) — the logic is backend-agnostic and lives in `gluten-core`.
### Bug description
`SparkResourceUtil.getTaskSlots` computes `executorCores / taskCores` with
no guard:
```scala
def getTaskSlots(conf: SparkConf): Int = {
val executorCores = SparkResourceUtil.getExecutorCores(conf)
val taskCores = conf.getInt("spark.task.cpus", 1)
executorCores / taskCores
}
```
`GlutenDriverPlugin.init` reads this slot count and divides by it (e.g.
`GlutenPlugin.setPredefinedConfigs` does `offHeapSize / taskSlots`), and four
other callers use it as a denominator too (`MemoryTargets`,
`ColumnarShuffleWriter`, and the Celeborn/Uniffle writers). Two invalid
configurations crash there with an opaque error instead of Spark's own clear
message:
- `spark.task.cpus > spark.executor.cores` (e.g. `executor.cores=1`,
`task.cpus=2`): integer division yields `0`, so a caller's `offHeapSize /
taskSlots` throws `ArithmeticException: / by zero`.
- `spark.task.cpus = 0`: `getTaskSlots` itself throws `ArithmeticException:
/ by zero`. `getTaskSlots` reads the value with raw `conf.getInt`, which
bypasses Spark's `CPUS_PER_TASK.checkValue(_ > 0)` (that check only fires on
the typed `conf.get(CPUS_PER_TASK)`).
The plugin runs at `SparkContext`'s `PluginContainer(this, ...)` step, which
is before `createTaskScheduler` and the resource-profile validation. So Gluten
throws first, and the user sees a Gluten `ArithmeticException` stack trace
instead of Spark's actionable messages (`validateTaskCpusLargeEnough`: "cores
per executor has to be >= cpus per task", or `CPUS_PER_TASK`: "should be
positive").
Both are invalid configs that Spark rejects anyway, so the job should not
start either way. The problem is the diagnostic: an opaque divide-by-zero with
a Gluten stack trace misattributes the failure to Gluten.
### Gluten version
main (1.8.0-SNAPSHOT)
### Spark version
Version-agnostic (applies to spark-3.3 / 3.4 / 3.5 / 4.0 / 4.1).
### Spark configurations
- `spark.plugins=org.apache.gluten.GlutenPlugin`
- Either `spark.task.cpus > spark.executor.cores`, or `spark.task.cpus=0`.
### System information
N/A — logic issue in
`gluten-core/src/main/scala/org/apache/spark/util/SparkResourceUtil.scala`,
independent of OS/hardware.
### Relevant logs
```text
java.lang.ArithmeticException: / by zero
at
org.apache.spark.util.SparkResourceUtil$.getTaskSlots(SparkResourceUtil.scala:81)
at
org.apache.gluten.GlutenPlugin$GlutenDriverPlugin.setPredefinedConfigs(GlutenPlugin.scala:...)
at org.apache.gluten.GlutenPlugin$GlutenDriverPlugin.init(...)
```
### Fix direction
Guard `getTaskSlots` so it never divides by a non-positive `taskCores` and
never returns 0: return a single slot for `taskCores <= 0`, and floor the
quotient at 1 otherwise. Gluten then defers to Spark's own validation for the
clear error message.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
