Phasic Energy: CFD simulations on AWS, run from the interface the engineers already knew

Overview

Phasic Energy designs thermal hardware: heat exchangers built for data centers, defense and heavy industry. Its engineers design with computational fluid dynamics, running large simulations in Siemens STAR-CCM+, and their models had grown past what a workstation could compute.

Cloudacio moved that work to AWS. A simulation can now use more than one machine, the engineers still drive it from the STAR-CCM+ interface they have always used, and the machines switch off when nobody is using them.

The Challenge

A CFD model that gets more detailed needs more cores and more memory, and at some point a single workstation runs out of both. The usual answer is to buy bigger hardware, and Phasic did not want to own and maintain a cluster that would sit unused between simulation campaigns.

Cloud compute solves the capacity problem and creates two others. The machines that run simulations of this size are expensive by the hour, so one left running over a weekend shows up on the bill. And each simulation produces files of tens of gigabytes that have to be kept, archived and protected.

There was also a condition that turned out to shape the whole design. Phasic’s engineers work inside the STAR-CCM+ graphical interface: they set a simulation up, watch it run and adjust it there. They were not willing to trade that for submitting batch jobs from a command line.

Our Solution

We started with a proof of concept, to put something real in front of the team quickly and learn from how they used it. A first version was running, and demoed, about 10 days after we got access to Phasic’s AWS account.

That first version followed the standard pattern for high-performance computing: a scheduler that queues simulations and runs them as jobs. Real use showed it did not fit. A queue assumes the engineer hands a job over and walks away, and Phasic’s engineers stay with the simulation while it runs. So we redesigned the platform around how they work.

STAR-CCM+ now runs on a workstation in AWS that engineers reach through a remote desktop, Amazon DCV, with the full graphical interface, as if it were under the desk. From that interface they send the heavy calculation to a dedicated high-performance node, so one simulation uses the cores and memory of more than one machine. The Siemens license server runs on AWS alongside it.

The rest runs without anyone thinking about it. Simulation files sync to Amazon S3, old intermediate files are cleaned up and results are archived for the long term. Machines stop on their own when they go idle, and we sized each instance type against what the workload actually used.

The Results

Phasic’s simulations are no longer limited by one machine. The first multi-node run used 178 parallel processes across two servers, launched from the same interface the engineers already knew.

Right-sizing cut the compute cost of the main server by about 75%. Its memory usage showed that a cheaper instance type was enough, at roughly a quarter of the price.

There is no hardware to buy or maintain. Capacity is switched on when a simulation needs it and off when it does not.

"Our models had outgrown our workstations, and we were not going to give up the way our engineers work to fix that. Cloudacio had STAR-CCM+ running on AWS in about ten days, and when a job queue turned out not to fit how we run simulations, they rebuilt the platform around us."Augie Smith, Founder at Phasic Energy