spark rdma error

[li7hui]

Hi,
I was trying to run SparkRDMA Terasort code. The common Spark Terasort can finish successfully, however, there exist errors for Spark RDMA Terasort code. Here is the errors as below:

I used Spark 2.1.0

[petro-rudenko]

Can you please provide command that you use to submit spark terasort? Do you use yarn or standalone cluster deployment mode? Can you please check logs for that blockManager Id?
Thanks,
Peter

[li7hui]

Hi,
I used the following commands to submit spark terasort job
./bin/spark-submit --class com.github.ehiggs.spark.terasort.TeraSort
--master spark:https://rdma21:7077 /root/spark-terasort-1.0-SNAPSHOT-jar-with-dependencies.jar
hdfs:https:///terasort_in hdfs:https:///terasort_out

I used the standalone mode.
I do not know where to get blockManagerId. I searched logs folder, but didn't find any blockManagerId there.

[petro-rudenko]

Ok can you please check for errors in spark log directory: grep -i error $SPARK_HOME/logs/. What dataset size do you run on?

[li7hui]

Hi, here are zip file in attachment which contains error message.

The dataset size I used is 1g.

spark-root-org.apache.spark.deploy.worker.Worker-1-rdma21.zip

[petro-rudenko]

Sorry logs for executors are in work directory.

[li7hui]

Hi,
Here is the error logs in work directory.
stderr.zip

[petro-rudenko]

In work directory there should be folders for each application and there separate folder for each executor. You need to collect executor logs from machines or use NFS. Tried to reproduce your case, it works for me:

1. Teragen 1g of data:

spark/bin/spark-submit -v --num-executors 10 --executor-cores 20 --executor-memory 24G --master yarn --class com.github.ehiggs.spark.terasort.TeraGen /hpc/scrap/users/peterr/spark-terasort/target/spark-terasort-1.1-SNAPSHOT-jar-with-dependencies.jar 1g /terasort-input-1g

2. Run terasort:

$ cat /hpc/scrap/users/swat/jenkins/spark/spark2.conf

spark.driver.extraJavaOptions   -Djava.library.path=/hpc/scrap/users/swat/jenkins/disni/
spark.executor.extraClassPath   /hpc/scrap/users/swat/jenkins//spark_rdma_artifacts/spark-rdma-2.0-for-spark-2.1.0-jar-with-dependencies.jar
spark.driver.extraClassPath     /hpc/scrap/users/swat/jenkins//spark_rdma_artifacts/spark-rdma-2.0-for-spark-2.1.0-jar-with-dependencies.jar
spark.executor.extraJavaOptions -Djava.library.path=/hpc/scrap/users/swat/jenkins/disni/
spark.shuffle.manager org.apache.spark.shuffle.rdma.RdmaShuffleManager
spark.executor.instances 16

$ bin/spark-submit -v --executor-cores 3 --properties-file /hpc/scrap/users/swat/jenkins/spark/spark2.conf --executor-memory 124G --master  spark:https://clx-orion-011:7077 --class com.github.ehiggs.spark.terasort.TeraSort /hpc/scrap/users/peterr/spark-terasort/target/spark-terasort-1.1-SNAPSHOT-jar-with-dependencies.jar /terasort-input-1g /terasort-output-1g

Could you please also try to generate bigger data. You are running 15 executors for 1 Gb of input data (<100Mb per executor). Or try to run with smaller number of executors to make sure everything is working

[li7hui]

Hi,
Thanks for your reply.
I used 4 nodes to start spark job, now the terasort spark job can start and finish successfully.
However, when the terasort spark rdma job running, my zabbix system can only see tcp traffic while the rdma traffic is zero. This is very strange! Because according to the stderr, the ibm.disni has been loaded during the processing.

[petro-rudenko]

Do you use Infiniband or Roce? Does your monitoring system configured to monitor RDMA traffic. You can check how to monitor RDMA traffic here: https://community.mellanox.com/docs/DOC-2416
Also you can run some ib perf tests from the perf package and make sure zabbix captures rdma traffic.

[li7hui]

Hi,
I use the RoCE network. I run some OFED perftest job,and the zabbix can see RMDA traffic.

What's your command to run SparkRDMA terasort please? I don't know whether there is some difference in the commands to run the SparkRDMA terasort program.

[petro-rudenko]

Here's how i run SparkRDMA Ehigg's terasort version:
Teragen 1g of data:

spark/bin/spark-submit -v --num-executors 10 --executor-cores 20 --executor-memory 24G --master yarn --class com.github.ehiggs.spark.terasort.TeraGen /hpc/scrap/users/peterr/spark-terasort/target/spark-terasort-1.1-SNAPSHOT-jar-with-dependencies.jar 1g /terasort-input-1g

Run terasort:

$ cat /hpc/scrap/users/swat/jenkins/spark/spark2.conf

spark.driver.extraJavaOptions   -Djava.library.path=/hpc/scrap/users/swat/jenkins/disni/
spark.executor.extraClassPath   /hpc/scrap/users/swat/jenkins//spark_rdma_artifacts/spark-rdma-2.0-for-spark-2.1.0-jar-with-dependencies.jar
spark.driver.extraClassPath     /hpc/scrap/users/swat/jenkins//spark_rdma_artifacts/spark-rdma-2.0-for-spark-2.1.0-jar-with-dependencies.jar
spark.executor.extraJavaOptions -Djava.library.path=/hpc/scrap/users/swat/jenkins/disni/
spark.shuffle.manager org.apache.spark.shuffle.rdma.RdmaShuffleManager
spark.executor.instances 16

$ bin/spark-submit -v --executor-cores 3 --properties-file /hpc/scrap/users/swat/jenkins/spark/spark2.conf --executor-memory 124G --master  spark:https://clx-orion-011:7077 --class com.github.ehiggs.spark.terasort.TeraSort /hpc/scrap/users/peterr/spark-terasort/target/spark-terasort-1.1-SNAPSHOT-jar-with-dependencies.jar /terasort-input-1g /terasort-output-1g

You can find how to run Hibench terasort version here, but the approach is the same. Basically you need to set spark.shuffle.manager org.apache.spark.shuffle.rdma.RdmaShuffleManager.

[li7hui]

Hi,
I think I found the problem. They set up PFC for the RoCE network in server and switch. The RoCE traffic will go under 5th queue, while other traffic,such as TCP, will go under 0th queue. Do you know how to setup disni to go under 5th queue?
Best,

[petro-rudenko]

Disni is just a wrapper over verbs api. If you setup PFC for rdma traffic to go under 5th queue, it'll go there. We've updated our wiki documentation, you can check Advanced forms of flowcontrol.
Let me know if you'll have questions. BTW we've released new version of SparkRDMA. Take a try ;) It has several performance improvements, bug fixes, more verbose error messages, etc.

[RummySugar]

我们的问题是：总是找不到libdisni，
19/08/28 16:57:44 ERROR RdmaNode: libdisni not found! It must be installed within the java.library.path on each Executor and Driver instance
我们反复检查了配置，找不出问题。能给我们一份您当时安装的详细步骤吗？

[tobegit3hub]

@RummySugar You need to install libdisni so in each server or upload with spark-submit.