[
https://issues.apache.org/jira/browse/SPARK-45527?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
wuyi updated SPARK-45527:
-------------------------
Description:
{code:java}
test("SPARK-XXX") {
import org.apache.spark.resource.{ResourceProfileBuilder,
TaskResourceRequests}
withTempDir { dir =>
val scriptPath = createTempScriptWithExpectedOutput(dir,
"gpuDiscoveryScript",
"""{"name": "gpu","addresses":["0"]}""")
val conf = new SparkConf()
.setAppName("test")
.setMaster("local-cluster[1, 12, 1024]")
.set("spark.executor.cores", "12")
conf.set(TASK_GPU_ID.amountConf, "0.08")
conf.set(WORKER_GPU_ID.amountConf, "1")
conf.set(WORKER_GPU_ID.discoveryScriptConf, scriptPath)
conf.set(EXECUTOR_GPU_ID.amountConf, "1")
sc = new SparkContext(conf)
val rdd = sc.range(0, 100, 1, 4)
var rdd1 = rdd.repartition(3)
val treqs = new TaskResourceRequests().cpus(1).resource("gpu", 1.0)
val rp = new ResourceProfileBuilder().require(treqs).build
rdd1 = rdd1.withResources(rp)
assert(rdd1.collect().size === 100)
}
} {code}
In the above test, the 3 tasks generated by rdd1 are expected to be executed in
sequence as we expect "new TaskResourceRequests().cpus(1).resource("gpu", 1.0)"
should override "conf.set(TASK_GPU_ID.amountConf, "0.08")". However, those 3
tasks are run in parallel in fact.
The root cause is that ExecutorData#ExecutorResourceInfo#numParts is static. In
this case, the "gpu.numParts" is initialized with 12 (1/0.08) and won't change
even if there's a new task resource request (e.g., resource("gpu", 1.0) in this
case). Thus, those 3 tasks are able to be executed in parallel.
was:
{code:java}
test("SPARK-XXX") {
import org.apache.spark.resource.{ResourceProfileBuilder,
TaskResourceRequests}
withTempDir { dir =>
val scriptPath = createTempScriptWithExpectedOutput(dir,
"gpuDiscoveryScript",
"""{"name": "gpu","addresses":["0"]}""")
val conf = new SparkConf()
.setAppName("test")
.setMaster("local-cluster[1, 12, 1024]")
.set("spark.executor.cores", "12")
conf.set(TASK_GPU_ID.amountConf, "0.08")
conf.set(WORKER_GPU_ID.amountConf, "1")
conf.set(WORKER_GPU_ID.discoveryScriptConf, scriptPath)
conf.set(EXECUTOR_GPU_ID.amountConf, "1")
sc = new SparkContext(conf)
val rdd = sc.range(0, 100, 1, 4)
var rdd1 = rdd.repartition(3)
val treqs = new TaskResourceRequests().cpus(1).resource("gpu", 1.0)
val rp = new ResourceProfileBuilder().require(treqs).build
rdd1 = rdd1.withResources(rp)
assert(rdd1.collect().size === 100)
}
} {code}
In the above test, the 3 tasks generated by rdd1 are expected to be executed in
sequence as we expect "new TaskResourceRequests().cpus(1).resource("gpu", 1.0)"
should override "conf.set(TASK_GPU_ID.amountConf, "0.08")". However, those 3
tasks are run in parallel in fact.
The root cause is that ExecutorData#ExecutorResourceInfo#numParts is static. In
this case, the "gpu.numParts" is initialized with 12 (1/0.08) and won't change
even if there's a new task resource request (e.g., resource("gpu", 1.0) in this
case). Thus, those 3 tasks are able to be executed in parallel.
> Task fraction resource request is not expected
> ----------------------------------------------
>
> Key: SPARK-45527
> URL: https://issues.apache.org/jira/browse/SPARK-45527
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
> Affects Versions: 3.2.1, 3.3.3, 3.4.1, 3.5.0
> Reporter: wuyi
> Priority: Major
>
>
> {code:java}
> test("SPARK-XXX") {
> import org.apache.spark.resource.{ResourceProfileBuilder,
> TaskResourceRequests}
> withTempDir { dir =>
> val scriptPath = createTempScriptWithExpectedOutput(dir,
> "gpuDiscoveryScript",
> """{"name": "gpu","addresses":["0"]}""")
> val conf = new SparkConf()
> .setAppName("test")
> .setMaster("local-cluster[1, 12, 1024]")
> .set("spark.executor.cores", "12")
> conf.set(TASK_GPU_ID.amountConf, "0.08")
> conf.set(WORKER_GPU_ID.amountConf, "1")
> conf.set(WORKER_GPU_ID.discoveryScriptConf, scriptPath)
> conf.set(EXECUTOR_GPU_ID.amountConf, "1")
> sc = new SparkContext(conf)
> val rdd = sc.range(0, 100, 1, 4)
> var rdd1 = rdd.repartition(3)
> val treqs = new TaskResourceRequests().cpus(1).resource("gpu", 1.0)
> val rp = new ResourceProfileBuilder().require(treqs).build
> rdd1 = rdd1.withResources(rp)
> assert(rdd1.collect().size === 100)
> }
> } {code}
> In the above test, the 3 tasks generated by rdd1 are expected to be executed
> in sequence as we expect "new TaskResourceRequests().cpus(1).resource("gpu",
> 1.0)" should override "conf.set(TASK_GPU_ID.amountConf, "0.08")". However,
> those 3 tasks are run in parallel in fact.
> The root cause is that ExecutorData#ExecutorResourceInfo#numParts is static.
> In this case, the "gpu.numParts" is initialized with 12 (1/0.08) and won't
> change even if there's a new task resource request (e.g., resource("gpu",
> 1.0) in this case). Thus, those 3 tasks are able to be executed in parallel.
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
