Film Studios Refocus HPC Management Lens
With the help of advanced supercomputing technologies, animated films have evolved from simple cycles of sketches to photorealistic, physics-accurate works of art.
As many of us have tracked, from the dawn of Pixar to the latest blockbusters where humans interact with realistic animations, there seems to be no limit to the hyper-realism in everything from a toss of hair, to a simulated breeze blowing snowflakes in real-time across a frozen tundra, to ships tossed by the laws of physics on rough, rendered oceans.
With great realism comes great investment, however. For instance, consider that from the time the Return of the King film emerged in 2003 with its stunning visuals until the more recent release of Dawn of the Planet of the Apes this year, there has been an estimated 10x increase in computational requirements, says David Beer, Senior Software Engineer at Adaptive Computing.
It’s no surprise that the advances in realism and accurate physics require an ever-mounting computational investment. While the media and entertainment segment represents only around an estimated, relatively steady 10% of the overall HPC server business, companies are looking to increase their compute—albeit in some interesting forms that fit the spotty, but demanding nature of the entertainment business.
The issue with rendering for many studios is that while the need for vast computational resources parallels those of large supercomputing sites, having a few thousand servers around permanently for only a relatively short period of work isn’t always efficient. Further, pushing this work out into the public cloud comes with its own concerns in terms of overall cost of provisioning large numbers of servers via special relationships with cloud service providers. Securing the servers is one side—on the cost front, the data movement adds significant weight to the bottom line while at top of mind for many studios are the still-lingering concerns about security.
Adding to all of this is the fact that most studios have conjured custom codes to manage the rendering of complex, character or scene-specific elements, which require custom workflows and management. In the end, says Beer, this means studios require a great deal of personalization, resource usage options, and workflow guidance along the way. While several studios have their own centrally-located datacenters (Pixar, for instance), others, including Weta Digital have their datacenter operations spread out over multiple centers. Still, the resources are nothing to overlook—when combined, these place in the high 100s on the Top 500.
When it comes to production studio rendering environments, the node count is often between 2,000 and 5,000 nodes, but the core counts can vary widely. Beer says that GPUs in particular are standard fare for the studios and while they might purchase a great many of them and take some advantage of mixed environments with boxes containing more memory or different coprocessors, their purchases tend to be more on the “vanilla” HPC side. In other words, studios aren’t experimenting with the latest HPC processor innovations—they’re sticking to what they know and focusing extensively on some other issues, including power and workload management.
While studios might be out front in terms of their use of GPUs and coprocessors, resource management remains a concern. While there are examples of production companies relying solely on both public or internal clouds for their rendering needs, Beer says they’re looking to make sure they’re maximizing investments in expensive HPC gear throughout the demand cycle—not just during times of peak operations. Accordingly, he says that companies like his own are finding ways to accommodate the needs of both larger studios that balance workloads across different sites as well as smaller media companies that can make use of the studio servers elsewhere via a private cloud arrangement. The point for them is that while it’s useful to house their own datacenters, for high demand for massive compute during large bursts of activity the ROI is clear, but during and after that wind-down, resource management becomes trickier.
“In the broader HPC world, they’ve solved a lot of the problems around utilization, efficiency and so forth. However, for the studios to see the utilization and power consumption rates that are found at the big supercomputing sites, there are still a lot of lessons to be learned,” Beer told us.
He also noted that as he talks with HPC community members about the concerns around exascale, the needs of the community devoted to that lofty goal and those that are driven by commercial enterprises, including media and entertainment, aren’t that different. In both realms, it’s about “getting your data where it’s supposed to be, having it ready for compute, and keeping the data pipeline stocked,” explained Beer.
“We are getting to a point when we build massive systems easily enough and while labs might have their concerns, for these people it’s worries about whether the power company will start charging big premiums. Exascale or commercial, it’s a dollars and sense conversation—but solving one solves the other.”
In short, whether it’s finding the right cloud balance, the right management framework, and the most efficient way to move and handle data, the scope of the systems for this industry aren’t out of the range of other HPC arenas. While the demand may be less ongoing, media and entertainment as a whole will be forced to face the need for growing computational requirements in increasingly complex ways.