IBM is actively working on adaptive features for the Map and Reduce phases of its InfoSphere BigInsights product (which is based on Apache Hadoop). In some cases, this involves applying techniques commonly found in mature data management products, and in some cases it involves developing new techniques. While a number of these adaptive features are still under development, there are some features in the product today. For instance, BigInsights currently includes an Adaptive Mapper capability that allows Mappers to successively process multiple splits for a job, and avoid the start-up costs for subsequent splits.

When a MapReduce job begins, Hadoop divides the data into multiple splits. It then creates Mapper tasks for each split. Hadoop deploys the first wave of Mapper tasks to the available processors. Then, as Mapper tasks complete, Hadoop deploys the next Mapper tasks in the queue to the available processors. However, each Mapper task has a start-up cost, and that start-up cost is repeated each time a Mapper task starts.

With BigInsights, there is not a separate Mapper task for each split. Instead, BigInsights creates Mapper tasks on each available processor, and those Mapper tasks successively process the splits. This means that BigInsights significantly reduces the Mapper start-up cost. You can see the results of a benchmark for a set-similarity join workload in the following chart. In this case, the tasks have a high start-up cost. The AM bar (Adaptive Mapper) in the chart is based on a 32MB split size. You can see that by avoiding the recurring start-up costs, you can significantly improve performance.

Of course, if you chose the largest split size (2GB), you would achieve similar results to the Adaptive Mapper. However, the you might potentially expose yourself to the imbalanced workloads that sometimes accompany very large splits.

The following chart shows the results of a benchmark for a join query on TERASORT records. Again the AM bar (Adaptive Mapper) in the chart is based on a 32MB split size.

In this case, the Adaptive Mapper results in a more modest performance improvement. Although, it is still an improvement. The key benefit of these Adaptive MapReduce features is that they eliminate some of the hassles associated with determining the split sizes, while also improving performance.

As I mentioned earlier in this post, a number of additional Adaptive MapReduce features are currently in development for future versions of BigInsights. I look forward to telling you about them when they are released…

In the mean time, make sure to check out the free online Hadoop courses at Big Data University. I previous blogged about my experiences with these courses in Hadoop Fundamentals Course on BigDataUniversity.com.

Here is a chart that compares the performance of Hadoop Distributed File System (HDFS) with General Parallel File System-Shared Nothing Cluster (GPFS-SNC) for certain Hadoop-based workloads (it comes from the Understanding Big Data book). As you can see, GPFS-SNC easily out-performs HDFS. In fact, the book claims that a 10-node GPFS-SNC-based Hadoop cluster can match the performance of a 16-node HDFS-based Hadoop cluster.

GPFS was developed by IBM in the 1990s for high-performance computing applications. It has been used in many of the world’s fastest computers (including Blue Gene and Watson). Recently, IBM extended GPFS to develop GPFS-SNC, which is suitable for Hadoop environments. A key difference between GPFS-SNC and HDFS is that GPFS-SNC is a kernel-level file system, whereas HDFS runs on top of the operating system. This means that GPFS-SNC offers several advantages over HDFS, including:

• Better performance
• Storage flexibility
• Concurrent read/write
• Improved security

If you are interested in seeing how GPFS-SNC performs in your Hadoop cluster, please contact IBM. Although GPFS-SNC is not in the current release of InfoSphere BigInsights (IBM’s Hadoop-based product), GPFS-SNC is currently available to select clients as a technology preview.