October 18, 2013
The Texas Advanced Computing Center (TACC) has been in the habit of spinning up some rather interesting machines these days, including the hybrid Stampede system. In early 2015, the center will be home to another notable resource—Wrangler, a data analysis and management cluster aimed at aiding the data-intensive need of the open science community.
Taking its place along Stampede’s side in the space that’s been left open from the retired Ranger machine, the new NSF-supported “big data” driven system will provide TACC and the communities it caters to with a Hadoop-ready Dell-supplied 120 node cluster. But that's not the real story here; what sets this apart is the anticipated high performance NAND flash side, supplied by the (still stealth) company, DSSD.
According to one of the PIs on the new system, Chris Jordan, their high performance NAND tier that is set to deliver one terabyte per second and a whopping 275 million IOPS.
The key to that kind of remarkable storage performance is coming from the technology provided by DSSD. Chances are, you haven’t heard of them before unless you follow news about the trajectory of Sun co-founder Andy Bechtolsheim’s career. His startup, DSSD, has been in development mode for well over three years and is the subject of a number of patents, although there are still no public customers or definitive products. As Jordan described when asked how they came to acquire their NAND flash products, their job at TACC is to keep an eye on emerging technologies and they’re “well connected” with Bechtolsheim and other companies on the edge of offering products publicly.
Of the patents in question from DSSD (there are three that could be found) one seems most promising (again, we note that TACC’s Chris Jordan was unable to give us any detail—this is speculation) there is one for a storage system with “guaranteed read latency” filed in 2012 from DSSD developed by William H. Moore, Jeffrey S. Bonwick. Here, they describe "A method for writing data to persistent storage. The method includes receiving a first request to write a first datum to persistent storage including NAND dies, identifying a first NAND die in which to write a first copy of the first datum and a second NAND die in which to write a second copy, generating a second request to write the first copy of the first datum to the first NAND die and a third request to write the second copy to the second NAND die, and waiting until the first NAND die and second NAND die not are busy. Based on a determination that the first NAND die and the second NAND die are not busy: issuing the second request to the first NAND die, and issuing the third request to the second NAND die after the second request is complete."
Again, we weren’t able to get any details, but we should note on a related front, Jordan says that the compute environment is what TACC defines as “embedded processing” which on a configuration level, is different than a typical Linux cluster setup with a large number of compute nodes and a separate storage subsystem with its own servers strung together with a high performance interconnect. Rather, in this case, storage will be closer to everything so that for the most part, users won’t go through an intermediate server to get to their data. This means fewer hops on the network between users and their data, which leads to higher performance and lower latency data access than what they might see with more horsepower-driven machine like Stampede.
Jordan tells us that Dell and DSSD are distinct, separate partners on the project and that while the NAND component wasn’t the sole basis for hardware decisions in general, it was a “very exciting part” of the initial concept. He noted that there are no special or custom Dell components for the system, but they did “work very closely” with Dell to achieve the desired result.
The Wrangler system will be the product of a $6 million NSF grant, which if you take some not-so-wild guesses, means that 120 nodes and some human support (the continuing support grant of another $6 million will be funded separately) equals quite a bit left over to fund this NAND storage effort.
Outside of the flashy side of the story, there are a few other elements worth noting. First, the system will be powered by 32 Haswell cores per node and while there are no hard, verified numbers to support the performance, we’ll be staying tuned to see how these early processors crunch some of the big data analytics problems the XSEDE and other scientific communities throw Haswell’s way. Further, to support the anticipated data-intensive workloads, they’ve made some noteworthy decisions on the memory front, adding 4 GB of RAM per core (versus 2 GB in a standard cluster) to lend an overall 128 GB of RAM to support faster storage access across the memory subsystem. Wrangler will also be able to rope in both 40 GbE and InfiniBand.
Additionally, this is one of a growing number of forays into the Hadoop and MapReduce space by a major research institution. TACC isn’t the first to install a Hadoop cluster, but according to Jordan, this cluster will likely grow—both in terms of additional nodes and the people required to support. Jordan told us that while at this point they’re using the native Apache Hadoop implementation, they haven’t ruled out the use of one of the commercial distributions (as offered by companies like Cloudera, MapR and Hortworks, for example).
Of the Hadoop, storage and processing environments, Jordan says that there were two real drivers for the design choices. First, he points to an increase in the overall need for a wider array of data analytics applications, which includes Hadoop and MapReduce type application, but also a host of other statistical and data mining tools as well as basic database applications. He says that while a traditional cluster environment can do all of those things, it’s far from optimal.
Additionally, he points to a growing class of persistent services for collecting, sharing and even analyzing data that are used by communities or large projects. These need to be available and accessible to cater to serve a cloud-based set of users. “Web users and web-based services are becoming a fundamental part of research in a way they haven’t been in the past,” he said, pointing to XSEDE and other projects, including domain-specific ones like iPlant, which serves as a science web application where users upload, share and analyze data or build their own VMs to run custom applications.
In addition to the system components we’ve already described, there will be two ten petabyte disk installations, one of which will be on site with the other at Indiana University, where it serve as an identical high capacity replicated storage resource.
We’ll catch up with TACC and hopefully DSSD at SC13 in Denver this year to see what we else we can learn.
In an earlier version of this article we referenced a comparison between the IOPS numbers of the TACC system with Blue Waters IOPS numbers that we derived from a Data Direct Networks statement. These were related to the storage subsystem and were not a valid reference for comparison. Notes from NCSA below..The article "Tiny TACC Cluster Set to Shatter IOPS Ceiling" included erroneous information about the Blue Waters system at NCSA.
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?