Configure application acceleration for an ECS instance

How it works

When you create an instance and deploy an application that supports performance acceleration, KeenTune performs full-stack performance tuning based on the application's workload characteristics. This tuning incorporates Alibaba Cloud's expertise to optimize not only the CPU, memory, I/O, and network, but also the application itself, ensuring your services run in an optimal performance environment. You can achieve an average performance improvement of about 20% for applications such as Nginx, MySQL, Redis, and PostgreSQL, which helps you reduce costs and improve efficiency. For details on the default application versions and their performance gains, see Default application versions and performance improvements.

Supported instances

The performance acceleration extension is supported only on select Yitian, AMD, and Intel instances that run Alibaba Cloud Linux 3.

New ECS instances

When you create a new ECS instance, you can install an application and configure performance acceleration during the creation process. After the instance is created, you can only disable performance acceleration or uninstall the application from the command line.

sudo bash /etc/keentune/target/scripts/set_xps_rps.sh eth0 rps disable
sudo keentune profile rollback
sudo systemctl stop keentune-target keentuned
sudo yum remove keentune-target keentuned

sudo systemctl stop <APP_Name>
sudo yum remove <APP_Name>

Existing ECS instances

If you have an existing Yitian instance (g8y, c8y, or r8y) that runs Alibaba Cloud Linux 3, you can install or uninstall the performance acceleration extension for applications on the instance details page.

• Supported performance acceleration extensions: x264/x265, Nginx, and Spark performance acceleration extension.

  Important

  • The Spark performance acceleration extension is in private preview. To use it, contact your account manager to enable the feature.

  • The Spark performance acceleration extension is supported only on ecs.g8y.8xlarge or higher instance types and ecs.r8y.8xlarge or higher instance types.

  • The Spark performance acceleration extension does not install Spark. It only installs and configures the tuning package. After you configure the extension, you must also modify the configuration file. For more information, see Configure the Spark performance acceleration extension.

  • The Spark performance acceleration extension supports only Spark 3.3. To request support for other versions, contact your account manager.

• For more information, see Install extensions and Uninstall extensions.

Default application versions and performance improvements

The following tables describe the default application versions and their performance improvements.

Configure the Spark performance acceleration extension

Follow these steps to configure the Spark performance acceleration extension on the Spark master node and all worker nodes.

1. Install the Spark performance acceleration extension on the Spark master node and all worker nodes.

   1. On the instance details page, choose Scheduled and Automated Tasks > Extension Installation/Uninstallation, and then click Install Extension.

   2. In the Install Extension dialog box, select Spark Performance Acceleration Extension from the Public Extension drop-down list, configure the following parameters, and then click Next. Follow the on-screen instructions to complete the process.

      • worker_number: The number of worker nodes in the Spark cluster.

      • worker_type: The instance type of the worker nodes. Currently, only ecs.g8y.8xlarge or higher instance types and ecs.r8y.8xlarge or higher instance types are supported.

2. Configure ZSTD on all Spark worker nodes.

   1. Replace the JAR files.

      for jar in $SPARK_HOME/jars/zstd-*.jar; do sudo mv "$jar" "${jar}.bak"; done
      sudo cp /opt/keentune/compress/zstd-*.jar $SPARK_HOME/jars/

   2. Modify the settings in the /opt/keentune/spark.conf file.

      • To configure the compression method for writing Parquet files, set the following parameters:

        spark.sql.parquet.compression.codec    zstd
        spark.hadoop.parquet.compression.codec.zstd.level 1

        Important

        If your Parquet version is earlier than 1.13, we recommend that you upgrade to 1.13 or later to enable the ZSTD buffer pool by default. For more information about the ZSTD JNI BufferPool, see Support ZSTD JNI BufferPool.

        wget https://repo1.maven.org/maven2/org/apache/parquet/parquet-column/1.13.1/parquet-column-1.13.1.jar
        wget https://repo1.maven.org/maven2/org/apache/parquet/parquet-common/1.13.1/parquet-common-1.13.1.jar
        wget https://repo1.maven.org/maven2/org/apache/parquet/parquet-encoding/1.13.1/parquet-encoding-1.13.1.jar
        wget https://repo1.maven.org/maven2/org/apache/parquet/parquet-format-structures/1.13.1/parquet-format-structures-1.13.1.jar
        wget https://repo1.maven.org/maven2/org/apache/parquet/parquet-hadoop/1.13.1/parquet-hadoop-1.13.1.jar
        wget https://repo1.maven.org/maven2/org/apache/parquet/parquet-jackson/1.13.1/parquet-jackson-1.13.1.jar
        for file in $SPARK_HOME/jars/parquet-*.jar; do sudo mv "$file" "$file.bak"; done
        sudo cp -rf parquet-*.jar $SPARK_HOME/jars

      • To configure the data compression method for the shuffle process, set the following parameters:

        spark.io.compression.zstd.level 1
        spark.io.compression.codec    zstd

3. If you use a separation of storage and compute architecture for Spark, you must configure the OSS endpoint and AccessKey.

   If you use OSS for storage, you must configure s3a-related parameters in the /opt/keentune/spark.conf file.

   spark.hadoop.fs.s3a.endpoint <Your_OSS_Endpoint>
   spark.hadoop.fs.s3a.access.key <Your_AccessKey_ID>
   spark.hadoop.fs.s3a.secret.key <Your_AccessKey_Secret>

4. On the Spark master node, start the Spark cluster by using the new configuration.

   You can start the Spark cluster by using one of the following methods:

   • Restart Spark by using the /opt/keentune/spark.conf file.

     # Method 1 for job submission:
     spark-submit --properties-file=/opt/keentune/spark.conf 
     # Method 2 for job submission:
     spark-sql --properties-file=/opt/keentune/spark.conf

   • Modify your own spark.conf file based on the optimized configuration, and then restart Spark.

     # Method 1 for job submission:
     spark-submit --properties-file=${your_spark.conf}
     # Method 2 for job submission:
     spark-sql --properties-file=${your_spark.conf}

Function

When you create an instance with the AI enhancement extension, KeenTune is installed. KeenTune optimizes OS resources such as the CPU, memory, and network based on the characteristics of AI training workloads. It also provides capabilities such as FUSE acceleration for CPFS. In certain AI framework training scenarios, this can improve performance by 15% or more. For details about specific frameworks, model versions, and performance benefits, see Frameworks, model versions, and performance.

Applicable instances

New ECS instances

You can install and configure performance acceleration when you create a new ECS instance. After the instance is created, you can uninstall the extension from the instance details page.

Uninstallation

On the instance details page, in the Installed Extensions list, click Detach in the Operation column.

Frameworks, models, and performance

We tested the performance of the OpenCLIP and BEVFormer models on mainstream GPU instances before and after optimization. The test environment and optimization results are provided below.

Performance results

The table below shows the performance improvement from FUSE acceleration for CPFS.

FUSE acceleration for CPFS is currently supported only on the Alibaba Cloud public image for Ubuntu 24.04. The following test data is based on this operating system version.