Technicolor speeds up movie production with Excelero NVMe flash

Media and entertainment giant Technicolor has gone live with 160TB of Excelero NVMe flash-based storage in a move that has allowed it to hugely boost performance in movie production processes.

Technicolor is headquartered in France and has more than 14,000 employees in 25 countries. But it was at its Los Angeles operations where teams worked on full-length motion pictures that Excelero was tested.

Technicolor is involved in the entire film production process from the set through to effects, editing and distribution. As a movie is worked on, it grows from around 10TB to around 350TB and the number of workstreams increases rapidly until there are a several users doing different things to the dataset.

This brings a need for concurrent access, which was previously handled – through Quantum StorNext and spinning disk HDD storage – by creating duplicate copies and/or scheduling access to the data.

The company wanted to solve the problem of concurrency without resorting to these measures, said Amir Bemanian, engineering director at Technicolor.

“We wanted enough performance for everyone to play in the same environment without the need for copying, duplicating and scheduling,” he said. “Until 2017 that had always been impossible, there was always a missing piece.”

Those missing pieces, said Bemanian, were hardware in the form of Intel’s NVMe flash drives and Excelero software to create LUNs on NVMe drives that could be served to its StorNext and GPFS file systems.

In a proof-of-concept started in December 2017, Technicolor deployed Excelero NVMesh software with 160TB on 40 Intel NVMe storage cards on SuperMicro commodity servers, with connectivity through a Mellanox 25/100Gbps Ethernet network.

NVMesh is aimed at webscale and cloud customers, large enterprises, analytics and high-performance computing-type environments. It is deployed at Nasa, where a 128-node cluster runs around 250TB of storage using its software for supercomputer simulations.

NVMesh allows servers to communicate with NVMe-equipped drives by NVMe-over-fabrics (NVMf) that uses remote direct access memory (RDMA) as a messaging layer for NVMe commands.

In terms of performance, Bemanian said it had surpassed expectations.

“The proof-of-concept was a real eye opener, with the bandwidth we were getting. We thought there was something wrong with the test parameters. Performance scaled linearly and we were getting near line speed with each stream,” he said.

“As we added streams, it didn’t degrade the performance of others, and we had 10 workstations with no degradation,” he added.

Throughput, said Bemanian, is now in the region of 48GBps. “Previously, we couldn’t deliver that from any of our file systems,” he said.

“We don’t have to manage performance. There’s no scheduling, no duplicating and we can play back at any resolution.”

Bemanian’s team hope to replace existing HDD storage entirely in 2019. Techncolor also hopes to benefit from improved TCO via Raid 6 support in Excelero, which currently only supports Raid 0, 1 and 10.

NVMe is the latest phase in the development of flash storage and is a subset of PCIe. It is a protocol written for flash storage that hugely increases access times and bandwidth compared with the SCSI and Sata protocols used by flash in “traditional” disk drive formats.

Right now, however, the industry is trying to find the best way to deploy it while keeping hold of advanced storage functionality provided by storage controllers.

Excelero’s approach is to put NVMe storage in servers and connect through NVMf and do away with the controller.

Other flash storage makers leave the controller in the way take a performance hit, while others aim to throw controller CPU resources at the problem through clustering, or through disaggregation of controller processing.