Purdue University, like most big research institutions, relies on a constantly evolving cyber infrastructure to support its large and diverse community of researchers. The diversity of workflows and exploding size of datasets have turned data storage and data management into major stumbling blocks in this environment. Recently, Purdue deployed a new 6.4 petabyte Data Depot, based on DataDirect Networks technology and achieved remarkable performance gains.
“We’ve been building some sort of research cyber infrastructure asset roughly annually since 2008,” said Mike Shuey, research infrastructure architect at Purdue. “At any time we have four to six mid-size and large scale cluster systems on the floor as well as five to ten smaller assets.” Some of these systems are purpose-built and may serve a particular instrument or even a single research project. The overall result is an extremely heterogeneous environment with multiple capabilities and heterogeneous workflows.
In fact, three of the university’s high-performance computing (HPC) systems appear on the current Top500 list, including the nation’s largest campus supercomputer (Conte), at number 73, and two other large-scale research clusters (Rice, at number 164, and Carter, at number 426.)
Coaxing optimum performance for the varying jobs is challenging. “The problem,” said Shuey, “is when you have larger data volumes and this level of heterogeneous assets, you’ll frequently see one project that wants to develop their data on one resource and analyze it on another.” For example, data collection from DNA sequencers works best using one cluster while data analysis requires a larger memory environment and runs better on a second cluster.
What’s needed “is a cross-cutting storage asset, a very capable file system, and a very capable data store that can be accessed concurrently at really high speeds by a dozen different instruments and compute systems and networks for several dozen different fields of science.” Importantly, said Shuey, this storage asset also must present directly to thousands of compute nodes – everything from a multimillion dollar cluster to a high-end Mac on someone’s desktop to a couple-year old embedded controller system that drives a scientific instrument.”
That’s a tall order. Purdue’s RFP went out last summer. Solutions from different vendors were evaluated, and DDN was the eventual winner. Purdue deployed a pair of DDN SFA12KX storage systems with SFX and 6.4 PBs of raw capacity for the university’s GPFS parallel file system. To ensure predictable, fast access to the Data Depot, Purdue also deployed DDN SFX Software to extend the storage cache with solid-state memory. As a result, the system loads the right data into flash storage at the right time to maximize cache hit rates and deliver a fast response.
An early decision that influenced other requirements was Purdue’s choice of IBM GPFS for the main file system.
“We have researchers pulling in data from instruments to a scratch file and this may be the sole repository of their data for several months while they are analyzing it, cleaning the data, and haven’t yet put it into archives. We are taking advantage of a couple of GPFS RAS (reliability, availability, and serviceability) features, specifically data replication and snapshot capabilities to protect against site-wide failure and to protect against accidental data deletion. While Lustre is great for other workloads – and we use it in some places – it doesn’t have those sorts of features right now,” said Shuey.
The storage also needed to be accessible by high-level policy management systems such iRODS and Globus, both of which are used in Purdue projects.
Important aspects of the DDN proposal included lower cost, technology flexibility, and high performance, according to Shuey: “DDN’s disk array, for example, lets us host solid state disks in the same enclosure so it reduced our overall component count. We don’t need separate disk enclosures and separate hardware for the metadata and the data. We can put it all in one asset and split it out logically within the DDN appliance.”
DDN’s relatively new SFX acceleration technology was also important. Again citing life science, which is an important part of Purdue’s research activity, Shuey noted LS workflows often have millions of tiny files whose IO access requirements can interfere with the more typical IO stream of simulation applications; larger files in a mechanical engineering simulation, for example, can be slowed by accesses to these millions of tiny files from a life sciences workflow.
“To avoid this, we’ve equipped SFX on all of our file system controllers so that basically any of those small reads will automatically get cached in a couple of terabytes of dedicated space,” said Shuey. “After a few months of operation we’re able to measure this and basically we’re getting about a 90 to 95 percent hit rate from the SFX cache, which means for read purposes we’re seeing almost a 900 percent improvement in our read capabilities.”
The result is faster performance and better isolation between user groups so that performance demands from one are less likely to disrupt others. “We get all of this without any real additional administrative overhead. This becomes a function that the storage system can deal with internally. It doesn’t require additional staffing or additional tuning, it just requires some dedicated hardware and a couple of licenses,” said Shuey.
The broad goal for the new Data Depot is to help Purdue to keep pace with its multidisciplinary research computational demands, including the surge in big data. For example, Purdue’s College of Agriculture recently teamed with the School of Mechanical Engineering to use sensor-equipped unmanned aircraft to collect critical data from acres of fields. New research outside the typical HPC realm also needed to be accommodated in the data repository, such as new projects from the College of Liberal Arts and Department of Sociology.
Rice, https://www.rcac.purdue.edu/compute/rice/