Below is an article by Nicole Hemsoth from HPCwire. A full version of this article can be found here
Purdue University made waves last year with its selection of HP’s POD containerized datacenter, which was hauled in to help them cope with a power inefficiencies stemming from an existing brick and mortar datacenter on campus.
The university set the proof point for cost and efficiency of modular datacenters, with their associate VP of Academic Technologies, John Campbell claiming that for 60% of the cost of a collocation facility the university could install a POD.
The selling point for containerized datacenters in general is that they come fully configured (although customizations can be made) with all the cables, power, cooling and racks in place and ready to roll. For Purdue, the savings mounted in the arenas of colo leasing, cutting back on staff to man datacenters, extension of on-campus networks, reduced power costs—which came, in part, because of the university’s own power plant.
UCLA announced this week that it has climbed aboard the containerized datacenter bandwagon with its head of academic technology services and managing director for the Institute of Digital Research and Education, Bill Labate, extolling the benefits of containerized HPC.
Labate’s group is responsible for providing university research cyberinfrastructure via its shared cluster system, which allows researchers who want to build their own clusters to instead buy compute nodes that Labate’s team integrates into the shared cluster. This allows the team to make the cycles available for over 170 research projects, from particle physics to genomcis and beyond.
As the need for cycles grew steadily, Labate saw a need for new equipment. He said that they had an existing datacenter that was a target for retrofitting, but when the team examined the possibility, it was clear there would be power and cooling limitations even though the space itself would have allowed room for growth. Labate’s team was able to secure $4.4 million to retrofit the existing data center, but when they received their final estimate for $7.2 million for the project, the shortfall led Labate down a different path.
Since it was not possible to scale down the potential retrofitted datacenter to remain within budget constraints, the possibilities of modular datacenters entered the picture. Labate said that to scale down to the level needed to suit the allotted funding would not have served even intermediate needs. Furthermore, since the goal of this undertaking was to enhance growth potential for the shared cluster resources, the retrofit would have been a waste of effort and money.
Labate approached UC San Diego for opinions about their experiences with a Sun-Oracle Black Box containerized solution, but found that they faced challenges with the U-shaped layout. UCSD told him that one thing they did not like was that the Black Box required specialized equipment and brought logistical challenges when it came to replacing and maintaining hardware since entire sections needed to be pulled out for fixes. This would not suit UCLA’s needs since, again, their system of buying new hardware was based on price-performance options among vendors, thus requiring flexibility to swap components based on what individual vendors offered. Besides, the Black Box solution was only a 20-foot container, and Labate knew that he needed to be able to power more cycles than the smaller Sun-Oracle solution could provide.
Labate’s team eventually settled on HP due to its high density, which was a good fit for what they were trying to accomplish in terms of providing as many cycles as possible. Other vendors they evaluated offered attractive density but Labate said there was not enough flexibility–that they needed to be able to grow with solutions that weren’t specialized for a particular container environment.
Before choosing the high-density, 40×8 feet POD container from HP, the team also looked at options from Dell, Rackable and as noted previously, the Sun-Oracle Black Box, which Labate says was the first to be struck from the list due to the size and shape limitations. He did not go into detail about the reasons behind abandoning the Dell and Rackable solutions, other than to say that for their specific needs, density was the deciding factor. Still, he noted that there were many similarities between the HP, Dell, IBM, and Rackable solutions—the choice simply came down to price, performance, flexibility of equipment solutions, and density.
The site preparations for the container began in October 2010 and moved swiftly until ending in mid-April of 2011. This entailed extending the university’s existing chilled water, power systems and pumps, fiber networks and laying the solid foundation required to support 110,00 pounds of steel and equipment.
Many modular datacenter makers emphasize the quick installation and set-up of their containers, claiming that it can be humming away in a few short weeks. As Labate says, however, anyone who knows anything about datacenters knows that you “can’t just plunk down a datacenter in your backyard and hook into your garden hose.” All told, from site prep to shared cluster bootup the team was looking at several months.
The shared cluster is distributed across campus with one building housing around 300 nodes, another with roughly 500 and now the POD, which packs in over 1500 nodes. His team ran a wide area InfiniBand network throughout, pulling all the nodes onto the same fabric for efficient management. They connected the Ethernet network for storage traffic, creating what he describes as a “geographic spread out single cluster.”
The team chose to keep the storage resources outside of the POD, in part to protect the valuable applications and results of long runs, but also because the POD has been optimized for compute nodes according to his team’s purpose to deliver shared cluster resources as if it was a single system. He emphasized repeatedly that their needs are specific—they wanted to be able to maximize the number of cycles available for university research.
When asked about usability or performance tradeoffs, Labate was adamant that containers are more efficient and perform for their needs, which again, are focused on providing more compute for the shared HPC cluster. He said that in many ways, the container streamlines their HPC operations by shedding the maintenance and efficiency hassles of brick and mortar. As he noted, “there are no other people in the POD, in fact, we limit our time in there since we want to keep it buttoned up as tight as possible. It’s been freeing, no operators in the pod, no need for anyone to sit in there and monitor—it’s all automated with all the tools we need for monitoring, powering on and off and so forth.”
According to Labate, there were no power and energy consumption problems with their use of POD. He said that compared to one of their brick and mortar datacenters which was operating at 1.5 PUE, the POD was running a steady 1.17 PUE. He claims that this translates into roughly a $200,000 difference in power costs, which represented a secondary but very important consideration as they looked at the POD capabilities.
Despite the lack of wide user adoption of modular datacenters, it was nearly impossible to get Labate to remark on any drawbacks to such solutions. He said that outside of the obvious negative factors, which include working inside small boxes with 36 raging blowers and tight quarters (which his team overcomes by saving fixes inside for once-weekly missions) and the aesthetic problem of having an giant, ugly shipping container fitting in with an artful sense of campus uniformity (an issue he said gave the campus aesthetics folk a few gripes) he can’t imagine traditional datacenters to address growth ever again.
When pressed about what he might warn others about when considering such solutions, Labate said environmental conditions were critical. First, in terms of making sure it is possible to locate the container close to needed power and cooling resources. Also, in terms of actually environment—he said that during a recent conversation with someone in an snow-bound region, he suggested that to avoid preventing access to the container they might need to consider building enclosures or renting indoor space.
Snow might not be a problem for UCLA, but earthquakes certainly are. Labate said this is another important distinction between brick and mortar and containers—while he notes he hasn’t researched his hunch, these massive, solid steel, windowless shipping containers were far likely more structurally sound than any existing traditional datacenter on his campus. Let’s hope he never gets a chance to prove that theory.