by clicking on the page. A slider will appear, allowing you to adjust your zoom level. Return to the original size by clicking on the page again.
the page around when zoomed in by dragging it.
the zoom using the slider on the top right.
by clicking on the zoomed-in page.
by entering text in the search field and click on "In This Issue" or "All Issues" to search the current issue or the archive of back issues respectively.
by clicking on thumbnails to select pages, and then press the print button.
this publication and page.
displays a table of sections with thumbnails and descriptions.
displays thumbnails of every page in the issue. Click on a page to jump.
allows you to browse through every available issue.
GCN : April 2015
24 GCN APRIL 2015 • GCN.COM STORAGE OPTIONS Flash memor y is more expensive than disk drives or tape, but it can be a good choice for performance-intensive ap- plications because of its speed. The DOE’s NERSC in Berkeley, Calif., deployed flash for its file system metadata in August. Hick said flash is working well in this key application, which affects all 6,000 of the center’s users for services such as logging in to the supercomputer. “Backing up our critical file system was taking 12 hours. Users noticed because the file system became slow and unusable,” Hick explained. “After deploying flash, the backups are down to three hours ... I’m not sure we have any complaints now. I don’t think the users even know we are backing it up.” Hick encourages data center opera- tors to conduct a careful analysis of flash memory and determine the trade- offs for each application. “There are certain cases where it makes a lot of sense to substitute flash even at an increased cost because the performance benefits are there,” Hick said. “If you can get four-times ben- efit in performance and reduce user complaints down to zero, I would say it’s worth it.” NERSC is so happy with how flash is working with its file system meta- data that it plans to have a layer of flash technology built inside of its next supercomputer. “The flash will be on [an] intercon- nect inside the supercomputer to store data for the duration of a simulation,” said Katie Antypas, deputy for data science at NERSC. “This will be another layer of storage that our users will have access to.” Antypas explained the Center’s different storage tiers: “ Now we have scratch, project and archive,” she said. “Scratch data we keep for up to 12 weeks. Project data we keep for a couple years, and our archive goes back 40 years. Flash will store data for hours or days. The trade-off is that it offers really high bandwidth.” John Goodhue, ex- ecutive director of the Massachusetts Green High Performance Computing Center (MGHPCC), has simi- lar views about the potential for flash. MGHPCC is a joint venture of five Massachusetts universities working on government-funded research projects such as climate change modeling, ge- nome sequencing and security analysis. The center uses flash for temporary storage when the supercomputer is working on a problem. “Flash gets you a lower latency. When you’re working with a dataset where you need to grab a large number of very small chunks of data out of a very big data set, flash is very good choice,” Goodhue said. “As the size of flash drives goes up and the cost goes down, the affordability of flash is improving over time.” Even so, Goodhue said sometimes regular disk drives are better than flash. “You need to think hard about the cost of flash and where it is really going to benefit you because it is very prob- lem-dependent,” Goodhue said. “Flash isn’t going to lift all boats, but it is going to lift a lot of them. Make sure that the speed really matters. Often, disk is just as good and is less expensive.” • The MGHPCC has adopted a cloud-based approach for the many petabytes of scientific data it stores in a two-year-old facility in Holyoke, Mass. MGHPCC has several tiers of stor- age. Scratch, for short-term data, provides 10 terabytes of temporary storage for computers that are work- ing on a problem. High-performance parallel file systems store petabytes of data after it has been processed and network-attached storage systems handle the most critical files, includ- ing home directories. “We use a cloud strategy for our storage,” Goodhue said. “That’s why we place a huge emphasis on high band- width and very efficient networking.” “What you’re seeing in a facility like ours is a large amount of data stored right next to the compute resources,” he said. “Instead of moving the data to the scientist’s computer, we’re moving the compute to where the large dataset is stored.” Goodhue said network speed is critical given that MGHPCC is located in Western Massachusetts while the researchers it supports are in Boston and other parts of the state. To work these digital distances, MGHPCC has 10G links to its university partners and plans to upgrade to 100G links. “We pride ourselves on looking like a local resource to our users,” Goodhue said. “It’s important from a networking and storage management point of view that it is very fast and very easy to move data from a workstation in Harvard out to Holyoke and back.” 1. CONSIDER FLASH 2. MAKE SURE THE NETWORK IS POWERFUL ENOUGH TO SUPPORT CLOUD STORAGE “There are certain cases where it makes a lot of sense to substitute flash even at an increased cost because the performance benefits are there.” — JASON HICK 0415gcn_022-027.indd 24 3/30/15 3:57 PM