The best performance for reading data into memory on the SAS LASR Analytic Server occurs when the server is co-located with the distributed data and the data is distributed evenly. The OLIPHANT procedure distributes the data such that parallel read performance by the SAS LASR Analytic Server is maximized. In addition, the distribution also ensures an even workload for query activity performed by the SAS LASR Analytic Server.

In order to produce an even distribution of data, it is important to understand that Hadoop stores data in blocks and that any block that contains data occupies the full size of the block on disk. The default block size is 32 megabytes and blocks are padded to reach the block size after the data is written. The data is distributed among the machines in the cluster in round-robin fashion. In order to maximize disk space, you can specify a block size that minimizes the padding.

It is important to know the size of the input data set such as the row count and the length of a row. This information, along with the number of machines in the cluster, can be used to set a block size that distributes the blocks evenly on the machines in the cluster and uses the space in the blocks efficiently.

For example, if the input data set is approximately 25 million rows with a row length of 1300 bytes, then the data set is approximately 30 gigabytes. If the hardware is a cluster of 16 machines, and 15 machines are used to provide HDFS storage, then storing 2 gigabytes on each machine is optimal. In this case, a BLOCKSIZE= setting of 32 megabytes or 64 megabytes would fill the overwhelming majority of blocks with data and reduce the space that is wasted by padding.

If the amount of data to add is not very large, then distributing it evenly can lead to poor block space utilization because at least one block is used on each machine in the cluster. However, the blocks might be mostly padding and contain little data. In these cases, the INNAMEONLY option can be used. This option sends the data to the Hadoop NameNode only. The blocks are distributed according to the default strategy used by Hadoop. The distribution is likely to be unbalanced, but the performance is not reduced because the data set is not large.