February 11, 2014

How HPC is Hacking Hadoop

Nicole Hemsoth

Although the trend may be quiet and distributed across only a relative few supercomputing sites, Hadoop and HPC are hopping hand-in-hand more frequently. These two technology areas aren’t necessarily made for another—there are limitations in what Hadoop can do. But a stretch of recent research has been pushing the possibilities, especially when it comes to making Hadoop fit data-intensive corners of scientific computing applications.

Despite the infrequency of news around Hadoop powering key research applications, we’ve watched key centers on this path, including the San Diego Supercomputer Center (which was one of the first to publish a comprehensive overview of using Hadoop on HPC resources) with great interest, and listened as nearly all major HPC system vendors (and many software ones too) targeted Hadoop users with key enhancements, tailored distributions, or even new product lines.

The research momentum behind Hadoop for HPC systems picked up in the last couple of years in particular. Among notable items are other explorations of Hadoop for data-intensive science, adapting MapReduce to an HPC environment, exploring it across different parallel file systems, handling scheduling and more. There are well over 2,000 peer-reviewed articles covering some aspect of this trend. The general theme, when you map it out and reduce it to a few words, is that the tooling required for HPC systems can be tweaked to fit Hadoop, especially when the purpose (offering a potential for more streamlined data management/processing on certain problems) is clear.

When it comes to data-intensive computing and Hadoop’s potential role in HPC, Dr. Glenn K. Lockwood at the San Diego Supercomputer Center (SDSC), is one of the key sources for information about specific challenges and opportunities. Most notably, Lockwood’s work on Hadoop for large-scale systems has drawn attention, particularly in terms of his work with the open source “big data” platform’s role on the Gordon system at SDSC.

Gordon is SDSC’s flash-based data-intensive computing resource. Although aimed at “big data” scientific computing, the Appro-built system still packs some serious compute power with its 16,160 cores, ranking at #88 on the most recent Top 500 list. The true measure of performance for Gordon, which was built to tackle data-intensive challenges, is in its input/output operations per second (IOPs) measurement—back when the machine was undergoing its acceptance cycle, it achieved 35 million IOPs. All of these elements made for some prime experimental ground for Lockwood and his colleagues.

In his role as a User Services Consultant at SDSC, Lockwood has been tracking a number of projects across the data-intensive computing spectrum. His most recent explorations, aside from running Hadoop clusters on Gordon, include writing Hadoop applications in Python with Hadoop Streaming, using (and finding parallel options for) the R language across supercomputers, and benchmarking several data-intensive computing frameworks, architectures and usage models.

“Although traditional supercomputers and Hadoop clusters are designed to solve very different problems and are consequentially architected differently, domain scientists are becoming increasingly interested in learning how Hadoop works and how it may be useful in addressing the data-intensive problems they face,” explained Lockwood.  “Making Hadoop available on Gordon has really made it easy for researchers to explore the features and benefits of Hadoop without having to learn an entirely new cloud API or be a systems administrator.”

He explained that instead, users can launch a Hadoop cluster by submitting a single pre-made job script to the batch system on Gordon with which they are already familiar. A “personal Hadoop cluster” is then launched on their job’s nodes, and users can then load data into their cluster’s distributed file system and run map/reduce tasks.  “Literally one single qsub command starts up a fully featured Hadoop cluster on Gordon’s 40 Gbps InfiniBand fabric with HDFS that is either backed by Gordon’s 300 GB SSDs or its Lustre filesystem,” said Lockwood. “This translates into Hadoop clusters capable of ingesting data to HDFS at a rate in excess of 750 MB/s and completing a 1.6 TB TeraSort in under 15 minutes. Because Gordon delivers this high performance both in traditional and Hadoop-based workloads, researchers can make meaningful performance comparisons on production-scale datasets.”

Lockwood highlighted how this has dramatically reduced the entry barrier for domain scientists who want to see what role Hadoop might play in their analyses, and it follows that training and exploratory work has driven a lot of the Hadoop use SDSC is currently seeing on Gordon.  “Faculty and researchers at universities nationwide have been using Gordon to teach courses in data analytics, and we’ve also been providing plenty of hands-on training to the local and national research communities via XSEDE, SDSC’s Summer Institute, PACE’s Data Mining Boot Camps, and UCSD’s Extension program.  In addition, we’ve provided cycles and classroom training for many applications built upon Hadoop including Mahout, Pig, HBase, and RHadoop.”

In Lockwood’s view, ultimately, Hadoop’s application in the traditional domain sciences is still in its infancy because the application ecosystem based on Hadoop is not as mature as the MPI-based ecosystem.  However, he says there is momentum in several non-traditional domains, including bioinformatics and anthropology, which are embracing Hadoop for production research on Gordon due to the natural fit of these domains’ problems with the map/reduce paradigm. “For example, we are supporting several projects that have begun exploring software built upon Hadoop such as Crossbow, CloudBurst, and SeqPig as scalable alternatives for massive genomic studies.  The evaluation process is still early on, but being able to run these Hadoop-based applications alongside the standard toolchain on Gordon is what is making the effort tractable.”

For anyone interested in the challenges and opportunities of deploying Hadoop on a complex system like Gordon, Lockwood has provided a rich overview here.

Aside from Lockwood’s work and that of his colleagues at SDSC, we wanted to point to some other projects that are helping HPC hack Hadoop to make it fit into a more complex environment. The following short list are a few of our top picks.

Tags: ,