Jeff Roach is always on the lookout for ways to improve the high-performance computations and data processing for scientific research at the University of North Carolina (UNC) at Chapel Hill. As Senior Scientific Research Associate in Information Technology Services/ Research Computing, he had been working with a team of Google engineers for a year when they introduced him to Techila Technologies, a Finnish startup that designed a computation engine that accelerates and automates data-intensive processing on Google Cloud . Rainer Wehkamp, Techila’s CEO, explains that the company partners with Google because “my engineers believe Google has the most technologically advanced cloud solutions.” And because Techila is already available through Google Cloud Marketplace, deployment was simple.

As a first project, Roach thought medical imaging would be a perfect fit: it requires large-scale computation and the ability to modify the calculations while in process. He set up a pilot with Wei-Tang Chang, Senior Research Associate in UNC’s Biomedical Research Imaging Center, who studies cortical layers of the brain through high-resolution functional magnetic resonance image (fMRI) datasets. Chang anticipated there would be benefits for researchers, clinicians, and patients: “While medical imaging with high spatial resolution provides more detailed information (i.e. higher sensitivity for diagnosis), it also increases the time that the patients need to stay in the scanner. By employing an imaging acceleration technique, we can reduce the scan time but patients and doctors need to wait longer for image reconstruction. With the computational acceleration of Techila and Google Cloud that issue could be greatly alleviated.”

Processing medical images in hours instead of weeks

Techila’s Distributed Computational Engine automatically scales and distributes computing across Google Compute Engine’s high-performance virtual machine (VM) instances. It works within platforms researchers are already familiar with, like PyCharm, MATLAB and R Studio, and lets them use their own cloud accounts. Tuomas Eerola, Techila’s Vice President of Business Development, says that they designed the computational engine for both speed and ease of use: “Techila adds abstraction that simplifies the technology for scientists. That gives them more time to do their analysis.”

"Roach and Chang experimented with running 40G of data with MATLAB on a local two-core node while on a second single-core node Techila’s engine distributed work to 100 nodes in the cloud on Google Cloud’s Compute Engine, each with four cores and 16G of memory. Roach reports that “the performance improvements gained on the proof-of-concept test case were outstanding. Depending on how busy we were, Chang used to get feedback in one to two weeks. With the Techila system and Google Cloud we were getting feedback in three hours. That’s a huge win.”

Chang agrees that the difference is dramatic: “Assuming the participant went through several functional runs and the total acquisition time is sixty minutes, the typical image reconstruction for one participant would take forty days, not even mentioning that the submitted jobs may queue on the servers. In other words, the image reconstructions for twenty participants in a study could take nearly two years! But with the computational acceleration of Techila's Compute Engine on Google Cloud, the image reconstruction that took more than a month before could be finished within eighteen hours. This enables quick feedback for optimization and troubleshooting.”

What’s next?

Soon Roach expects to expand this pilot to UNC researchers in genetics and computer science—and he’s eager to try running it on Google Cloud’s Pre-emptible VMs, which can lower computing costs by up to 80%. Eerola explains that Techila was designed for automation so it has built-in flexibility to adapt and shift resources when VMs are pre-empted. That enables researchers to benefit from the lower costs of Pre-emptible VMs without interrupting their computations. In the meantime, the pilot has already advanced Roach and Chang’s original goals: to speed up the time to discovery so researchers could ask more questions of the data, pose more hypotheses, and ultimately get answers faster.