commit | 0c826c26a547597b9f97e24c809ff073c2ffd30b | [log] [tgz] |
---|---|---|
author | John Yang <johnyangk@gmail.com> | Tue Oct 30 10:13:18 2018 +0900 |
committer | Jangho Seo <jangho@jangho.io> | Tue Oct 30 10:13:18 2018 +0900 |
tree | c8c26d70382c3e4854a7736ce93675b00d46bdfc | |
parent | 970751acfb4f9ca357310db6d12f1ccd6841cc43 [diff] |
[NEMO-8] Implement PipeManagerMaster/Worker (#129) JIRA: [NEMO-8: Implement PipeManagerMaster/Worker](https://issues.apache.org/jira/projects/NEMO/issues/NEMO-8) **Major changes:** - Supports fully-pipelined data streaming for bounded sources (not unbounded sources) - Tasks do 'finish' after processing all input data, as the data is finite - When a tasks finishes, it emits all data it has (e.g., GroupByKey accumulated results) and closes corresponding outgoing pipes, notifying downstream tasks the end of the pipes - For stream-processing unbounded sources, we need watermarks (https://issues.apache.org/jira/browse/NEMO-233) - Introduces PipeManagerMaster/Worker - Shares code with BlockManagerMaster/Worker - Naive, Element-wise serialization+compression+writeAndFlush - Very likely that this will cause some serious overheads. Will run proper benchmarks and fix the issues in a later PR. **Minor changes to note:** - JobConf#SchedulerImplClassName: Batch and Streaming options - StreamingPolicyParallelismFive: The default policy + PipeTransferEverythingPass - Fixes the StreamingScheduler to pass the new streaming integration tests - Fixes a coder bug in the Beam frontend (PCollectionView coder) **Tests for the changes:** - WindowedWordCountITCase#testStreamingFixedWindow - WindowedWordCountITCase#testStreamingSlidingWindow **Other comments:** - Also closes "Implement common API for data transfer" (https://issues.apache.org/jira/browse/NEMO-9) Closes #129
A Data Processing System for Flexible Employment With Different Deployment Characteristics.
Details about Nemo and its development can be found in:
Please refer to the Contribution guideline to contribute to our project.
export HADOOP_HOME=/path/to/hadoop-2.7.2 export YARN_HOME=$HADOOP_HOME export PATH=$PATH:$HADOOP_HOME/bin
On Ubuntu 14.04 LTS and its point releases:
sudo apt-get install protobuf-compiler
On Ubuntu 16.04 LTS and its point releases:
sudo add-apt-repository ppa:snuspl/protobuf-250 sudo apt update sudo apt install protobuf-compiler=2.5.0-9xenial1
On macOS:
brew tap homebrew/versions brew install protobuf@2.5
Or build from source:
./configure
make
make check
sudo make install
To check for a successful installation of version 2.5.0, run protoc --version
mvn clean install -T 2C
mvn clean install -DskipITs -T 2C
-job_id
: ID of the Beam job-user_main
: Canonical name of the Beam application-user_args
: Arguments that the Beam application accepts-optimization_policy
: Canonical name of the optimization policy to apply to a job DAG in Nemo Compiler-deploy_mode
: yarn
is supported(default value is local
)## MapReduce example ./bin/run_beam.sh \ -job_id mr_default \ -executor_json `pwd`/examples/resources/beam_test_executor_resources.json \ -optimization_policy org.apache.nemo.compiler.optimizer.policy.DefaultPolicy \ -user_main org.apache.nemo.examples.beam.WordCount \ -user_args "`pwd`/examples/resources/test_input_wordcount `pwd`/examples/resources/test_output_wordcount" ## YARN cluster example ./bin/run_beam.sh \ -deploy_mode yarn \ -job_id mr_transient \ -executor_json `pwd`/examples/resources/beam_test_executor_resources.json \ -user_main org.apache.nemo.examples.beam.WordCount \ -optimization_policy org.apache.nemo.compiler.optimizer.policy.TransientResourcePolicy \ -user_args "hdfs://v-m:9000/test_input_wordcount hdfs://v-m:9000/test_output_wordcount"
-executor_json
command line option can be used to provide a path to the JSON file that describes resource configuration for executors. Its default value is config/default.json
, which initializes one of each Transient
, Reserved
, and Compute
executor, each of which has one core and 1024MB memory.
num
(optional): Number of containers. Default value is 1type
: Three container types are supported:Transient
: Containers that store eviction-prone resources. When batch jobs use idle resources in Transient
containers, they can be arbitrarily evicted when latency-critical jobs attempt to use the resources.Reserved
: Containers that store eviction-free resources. Reserved
containers are used to reliably store intermediate data which have high eviction cost.Compute
: Containers that are mainly used for computation.memory_mb
: Memory size in MBcapacity
: Number of Task
s that can be run in an executor. Set this value to be the same as the number of CPU cores of the container.[ { "num": 12, "type": "Transient", "memory_mb": 1024, "capacity": 4 }, { "type": "Reserved", "memory_mb": 1024, "capacity": 2 } ]
This example configuration specifies
Nemo Compiler and Engine can store JSON representation of intermediate DAGs.
-dag_dir
command line option is used to specify the directory where the JSON files are stored. The default directory is ./dag
. Using our online visualizer, you can easily visualize a DAG. Just drop the JSON file of the DAG as an input to it../bin/run_beam.sh \ -job_id als \ -executor_json `pwd`/examples/resources/beam_test_executor_resources.json \ -user_main org.apache.nemo.examples.beam.AlternatingLeastSquare \ -optimization_policy org.apache.nemo.compiler.optimizer.policy.TransientResourcePolicy \ -dag_dir "./dag/als" \ -user_args "`pwd`/examples/resources/test_input_als 10 3"