How Does the SAS LASR Analytic Server Work?

Distributed SAS LASR Analytic Server

The server provides a client/server environment where the client connects to the server, sends requests to the server, and receives results back from the server. The server-side environment is a distributed computing environment. A typical deployment is to use a series of blades in a cluster. In addition to using a homogeneous hardware profile, the software installation is also homogeneous. The same operating system is used throughout and the same SAS software is installed on each blade that is used for the server. In order for the software on each blade to share the workload and still act as a single server, the SAS software that is installed on each blade implements the Message Passing Interface (MPI). The MPI implementation is used to enable communication between the blades.

After a client connection is authenticated, the server performs the operations requested by the client. Any request (for example, a request for summary statistics) that is authorized will execute. After the server completes the request, there is no trace of the request. Every client request is executed in parallel at extraordinarily high speeds, and client communication with the server is practically instantaneous and seamless.

There are two ways to load data into a distributed server:

load data from tables and data sets. You can start a server instance and directly load tables into the server by using the SAS LASR Analytic Server engine or the LASR procedure from a SAS session that has a network connection to the cluster. Any data source that can be accessed with a SAS engine can be loaded into memory. The data is transferred to the root node of the server and the root node distributes the data to the worker nodes. You can also append rows to an in-memory table with the SAS LASR Analytic Server engine.
load tables from a co-located data provider.
- Tables can be read from the Hadoop Distributed File System (HDFS). You can use the SAS Data in HDFS engine to add tables to HDFS. When a table is added to HDFS, it is divided into blocks that are distributed across the machines in the cluster. The server software is designed to read data in parallel from HDFS. When used to read data from HDFS, the LASR procedure causes the worker nodes to read the blocks of data that are local to the machine.
- Tables can also be read from a third-party vendor database. For distributed databases like Teradata and Greenplum, the SAS LASR Analytic Server can access the local data on each machine that is used for the database.

The following figure shows the relationship of the root node, the worker nodes, and how they interact when working with large data sets in HDFS. As described in the previous list, the LASR procedure communicates with the root node and the root node directs the worker nodes to read data in parallel from HDFS. The figure also indicates how the SAS Data in HDFS engine is used to transfer data to HDFS.

Relationship of PROC LASR and the SAS Data in HDFS Engine

Note: The preceding figure shows a distributed architecture that uses HDFS. For deployments that use a third-party vendor database, the architecture is also distributed, but different procedures and software components are used for distributing and reading the data.

After the data is loaded into memory on the server, it resides in memory until the table is unloaded or the server terminates. After the table is in memory, client applications that are authorized to access the table can send requests to the server and receive the results from the server.

In-memory tables can be saved. You can use the SAS LASR Analytic Server engine to save an in-memory table as a SAS data set or as any other output that a SAS engine can use. This method of using an engine transfers the data across the network connection. For large tables, saving to HDFS is supported with the LASR and IMSTAT procedures. This strategy saves the data in parallel and keeps the data on the cluster.

Non-Distributed SAS LASR Analytic Server

Most of the features that are available with a distributed deployment also apply to the non-distributed deployment too. Any limitations are related to the reduced functionality of using a single-machine rather than a distributed computing environment.

In a non-distributed deployment, the server acts in a client/server fashion where the client sends requests to the server and receives results back. The server performs the analytic operations on the tables that are loaded in to memory. As a result, the processing times are very fast and the results are delivered almost instantaneously.

You can load tables to a non-distributed server with the SAS LASR Analytic Server engine. Any data source that SAS can access can be used for input and the SAS LASR Analytic Server engine can store the data as an in-memory table. The engine also supports appending data.

You can save in-memory tables by using the SAS LASR Analytic Server engine. The tables can be saved as a SAS data set or as any other output that a SAS engine can use.