It isn’t often that you can get access to a thousand-node network to test your latest app, but thanks to the efforts of EMC’s Greenplum unit and some additional computing vendors, you can, and more amazingly, it is free of charge too.
The network was announced last fall at Strata and connects 1,000 specialized servers from Supermicro running dual Intel Xeon processors with 48 GB of RAM apiece along with Mellanox 10 GB Ethernet adapters and switches, and a total of 12,000 Seagate 2 TB drives. It is all contained within Greenplum’s Las Vegas data center, with the goal of having the largest publically accessible Hadoop cluster around. While Yahoo and eBay and others have some fairly large Hadoop clusters, they generally don’t let anyone else come in and try out their apps. The cluster goes under the name of Analytics Workbench. On this page, you can click on the “learn more” button and submit your name if you are interested in using the cluster.
The goal, according to Greenplum staffers, is to have a community and collaborative big data platform that can be applied to a set of analytical problems that have wide appeal. When the Strata announcement was made last fall, Greenplum stated that they wanted to eventually publish any results from the cluster, but they haven’t yet. Intel was one of the first clients to use the workbench (and running a thousand-node job too), but they are still reviewing their results.
Other clients that are running tests on the cluster include Mellanox and VMware, who both donated gear to power it, and a research team from the University of Central Florida. A group from NASA Goddard is using it to perform an analysis of historical weather patterns. The cluster formally opened up in July, and yes, it is really is free of charge. Applicants need to be vetted and work closely with the Greenplum engineers to get their apps uploaded and configured to the cluster.
“We accept bids based on any submitted application and developers can request specific time and resources,” says William Davis, one of the Greenplum product marketers involved with the cluster’s creation. Applications are reviewed by an internal group of Hadoop experts called the Jedi Council, and they try to select who will have the best fit for the next test run on the cluster.
Greenplum intends to use the cluster in a variety of ways besides public testing. Sometime next quarter they will launch a training program for Hadoop. A unique aspect of the program is that each member of the course will be granted access to the cluster to use as a sandbox environment for their own project. They are still working out the details on how this will work. The company has other fee-based programs to leverage its experience with this cluster, including what it calls its Analytics Lab packages. This uses their team of data scientists on specific vertical markets or particular custom applications.
There are several other tools that are offered on the cluster in addition to Hadoop including MapReduce, the parallel job processing software; VMware’s Rubicon system management team; and standard Hadoop add-ons such as Hive, Pig, and Mahout.
Greenplum isn’t the first to have such a large test bed assembled, but probably the first to use this level of gear for Hadoop and other data science activities. In the late 1980s, a group of Novell engineers in Utah created the “SuperLab” which eventually grew to1,700 PCs connected together. The lab was used to prove the features and scalability of Novell’s Netware network operating system, a piece of software that at one time could be found in most enterprises but now is largely a historical curiosity. Just to give you some perspective, in 1999 the PCs in Novell’s lab had a whopping 256 MB of RAM and 8 GB of storage (try buying that on today’s PCs). How times have changed.
Anyway, the SuperLab team left Novell a few years later and built their own private test lab for a startup called Keylabs. I was one of their early customers, using the facility to publish some of the test results in cNet and other IT publications of the first Web server comparison tests.
The Keylabs engineers very quickly discovered that automating the sequencing and actions of the individual PCs was tedious, and they wrote software that eventually spawned Altiris. Part of the assets of this company was later purchased by Symantec and is still used for their desktop imaging and management tool line.
Speaking of scaling up to a thousand machines automatically, running tests on this scale can be tricky. Greenplum has already seen several hardware failures that take down particular nodes as they have begun using their cluster. And like Keylabs, understanding how to sequence all this gear to come online quickly can be vexing: imagine if each machine takes just ten minutes to boot up and launch an app: times ten or twenty nodes that isn’t much of a big deal, but when you are trying to bring up hundreds it could tie up the cluster for the better part of a week in just starting up the tests. “It is a bit of a challenge in educating our customers on how to use and manage something of this size and how to deploy their software across the entire cluster. You can’t deploy software serially, and we have to make sure that our customers understand these issues,” says Davis.
So get your application in now for testing your app. You could be making computing history.