|
| 1 | +# SparkRDMA Shuffle Manager Plugin |
| 2 | +The SparkRDMA Plugin is a high performance shuffle manager that uses RDMA (instead of TCP) when |
| 3 | +performing the shuffle phase of the Spark job. |
| 4 | + |
| 5 | +This open-source project is developed, maintained and supported by [Mellanox Technologies](http://www.mellanox.com). |
| 6 | + |
| 7 | +## Performance results |
| 8 | +Example performance speedup for HiBench TeraSort: |
| 9 | + |
| 10 | + |
| 11 | +Running TeraSort with SparkRDMA is x1.41 faster than standard Spark (runtime in seconds) |
| 12 | + |
| 13 | +Testbed: |
| 14 | + |
| 15 | +175GB Workload |
| 16 | + |
| 17 | +15 Workers, 2x Intel Xeon E5-2697 v3 @ 2.60GHz, 28 cores per Worker, 256GB RAM, non-flash storage (HDD) |
| 18 | + |
| 19 | +Mellanox ConnectX-4 network adapter with 100GbE RoCE fabric, connected with a Mellanox Spectrum switch |
| 20 | + |
| 21 | +## Runtime requirements |
| 22 | +* Apache Spark 2.0.0 (more versions to be supported) |
| 23 | +* Java 8 |
| 24 | +* libdisni 1.2 |
| 25 | +* An RDMA-supported network, e.g. RoCE or Infiniband |
| 26 | + |
| 27 | +## Build |
| 28 | + |
| 29 | +* Building the SparkRDMA plugin requires [Apache Maven](http://maven.apache.org/) and Java 8 |
| 30 | + |
| 31 | +1. Obtain a clone of [SparkRDMA](https://github.com/Mellanox/SparkRDMA) |
| 32 | + |
| 33 | +2. Build the plugin: |
| 34 | +``` |
| 35 | +mvn -DskipTests package |
| 36 | +``` |
| 37 | + |
| 38 | +3. Obtain a clone of [DiSNI](https://github.com/zrlio/disni) for building libdisni 1.2: |
| 39 | + |
| 40 | +``` |
| 41 | +git clone https://github.com/zrlio/disni.git |
| 42 | +cd disni |
| 43 | +git checkout -b v1.2 247fe8abe54c90b450d2a4b0679e59cfa83205f6 |
| 44 | +``` |
| 45 | + |
| 46 | +4. Compile and install only libdisni (the jars are already included in the SparkRDMA plugin): |
| 47 | + |
| 48 | +``` |
| 49 | +cd libdisni |
| 50 | +sh autoprepare.sh |
| 51 | +./configure --with-jdk=/path/to/java8/jdk |
| 52 | +make |
| 53 | +make install |
| 54 | +``` |
| 55 | +5. libdisni **must** be installed on every Spark Master and Worker |
| 56 | + |
| 57 | +## Configuration |
| 58 | + |
| 59 | +* Provide Spark the location of the SparkRDMA plugin jars by using the extraClassPath option. For standalone mode this can |
| 60 | +be added to either spark-defaults.conf or any runtime configuration file. For client mode this **must** be added to spark-defaults.conf |
| 61 | + |
| 62 | +``` |
| 63 | +spark.driver.extraClassPath /path/to/SparkRDMA/target/spark-rdma-1.0-jar-with-dependencies.jar |
| 64 | +spark.executor.extraClassPath /path/to/SparkRDMA/target/spark-rdma-1.0-jar-with-dependencies.jar |
| 65 | +``` |
| 66 | + |
| 67 | +## Running |
| 68 | + |
| 69 | +* To enable and use the SparkRDMA Shuffle Manager plugin, add the following line to either spark-defaults.conf or any runtime configuration file: |
| 70 | + |
| 71 | +``` |
| 72 | +spark.shuffle.manager org.apache.spark.shuffle.rdma.RdmaShuffleManager |
| 73 | +``` |
| 74 | + |
| 75 | +## Contributions |
| 76 | + |
| 77 | +Any PR submissions are welcome |
0 commit comments