USENIX researchers get a grip on Hadoop performance
Now that big data technologies like Apache Hadoop are moving into the enterprise, system engineers must start building models that can estimate how much work these distributed data processing systems can do and how quickly they can get their work done.
Having accurate models of big data workloads means organizations can better plan and allocate resources to these jobs, and can confidently assert when the results of this work can be delivered to customers.
Estimating big data jobs, however, is tricky business, and the process cannot rely entirely on traditional modeling tools, according to researchers speaking at the USENIX annual conference on autonomic computing, being held this week in Philadelphia.
“It’s almost impossible to be accurate, because you are dealing with a non-deterministic system,” said Lucy Cherkasova, a researcher at Hewlett-Packard Labs.
She explained that Hadoop systems are non-deterministic because they have a wide range of variable factors that can contribute to how long it takes for a job to finish.
The average Hadoop system might have up to 190 parameters to set in order to start running, and each Hadoop job may have different requirements for how much computation, bandwidth, memory or other resources it needs.
Cherkasova has been working on models, and associated tools, to estimate how long a large data processing job will take to run on Hadoop or other large data processing systems, in a project called ARIA (Automatic Resource Inference and Allocation for MapReduce Environments).
ARIA aims to answer the question, “How many resources should I allocate to this job, if I want to process this data by this deadline,” Cherkasova said.
One might assume that if you double the number of resources of a Hadoop job, the time required to complete the job would be cut in half. “This is not the case” with Hadoop, Cherkasova said.
Job profiles can change in non-linear ways depending on the number of servers being used. The performance bottlenecks in a Hadoop cluster for 66 nodes are different from the bottlenecks found in a Hadoop cluster of 1,000 nodes, she said.
The performance can vary according to the type of job as well. Some of the research Cherkasova carried out involved studying what sized virtual machine would be best suited for Hadoop jobs.
For instance, Amazon Web Services (AWS) offers a range of virtual servers, from small instances with a single processor to larger ones with eight or more processors. Because Hadoop is a distributed system, it was made to run on multiple servers. But would it be more cost-effective to run Hadoop across many smaller instances, or on fewer though larger smaller instances?
Cherkasova found that the answer depends on the workload.
One type of job, http://www.highlyscalablesystems.com/3235/hadoop-terasort-benchmark/”>Terasort, in which a large amount of data is sorted, can be completed five times more quickly by using a collection of small AWS instances compared to using the large instances.
The performance of another type of job, the Kmeans clustering algorithm, does not vary with the kind of instance used, however. It runs equally well on small, medium, or large instances, meaning the user can run a Kmeans job on the more cost-effective large instances without sacrificing any speed.
Cherkasova’s work in this field has been important because to date there have been very few widely cited studies on modeling Hadoop performance, said Anshul Gandhi, an IBM researcher who was on the USENIX organizing committee for the conference.
Studying Hadoop can be a challenge because few researchers have access to large Hadoop systems, which are too costly to build and test, Gandhi said.
Also doing work in this realm has been Cristina Abad, a computer science Ph.D. candidate at the University of Illinois at Urbana-Champaign.
Abad has developed a benchmark designed to model the performance of next-generation storage systems, called MimesisBench, and has modeled a workload on a Yahoo 4,100 node cluster running on the Hadoop Distributed File System (HDFS).
The benchmark can help determine if a storage system can accommodate an increased workload, which can be valuable information for determining whether to make major architectural changes when increasing the throughput of a data processing system.
The benchmark showed, for instance, that the Yahoo cluster would start experiencing increased latency when handling approximately more than 16,800 operations per second, which was greater than was expected.
The benchmark could also help in other architectural decisions. For its storage system, Yahoo used a hierarchal namespace, in which files are organized into groups or subdirectories. If Yahoo were to use a flat namespace, where all the files are located in a single list, latency would have started spiking at about 10,284 operations per second, the model showed.