Cycle Computing and the HPC Experiment
With hardware advancing at a relatively stable (if still exponential) rate and datasets increasing at a much higher rate, parallelism is a main tenet of high performance computing today. That parallelism is difficult to attain in a cloud environment, as latencies there are typically higher, thus slowing performance.
Three weeks ago, Jason Stowe, CEO of Cycle Computing, spoke with HPC in the Cloud about their work in renting large clusters of Amazon HPC instances for companies looking for a short but powerful burst of that parallelized computing power. The focus was on how they aided Schrodinger in winning a Bio-IT Best Practices award with their intensive yet relatively inexpensive protein calculations.
That conversation took place on the heels of a presentation done with Wolfgang Gentzsch and Burak Yenier in association with the HPC Experiment, where Stowe went more into detail about the additional use cases in which Cycle Computing has facilitated HPC experimentation in the cloud. Gentzsch and Yenier also went on to provide an update on the HPC experiment, the fourth round of which kicks off this week.
The problem with in-house HPC equipment, according to Stowe, lies in a lack of stability in resource requirement. Oftentimes the HPC cluster goes under-utilized, meaning relatively expensive machinery is idling on valuable floor space. On the other end of the spectrum, the servers may not fulfill the peak needs of the institution.
“The clusters are too small when you need them most…you generally wish it was several times larger than it actually is,” Stowe said of companies with in-house resources when they face the peak of their intermittent computing needs schedule.
The challenge was based on the knowledge that, according to Stowe, that some top ten pharmaceutical companies can run approximately 341,700 hours of computing against a cancer target every year.
“We essentially were able to run very large sets of compounds an order of magnitude over what they normally would have been able to provision against different cancer affiliated proteins,” he said.
As Stowe noted, their work was in drug design and running simulations on how to either stimulate or halt protein activity. “What you’re trying to do,” he said, “is knock small molecules that lock so that they either inhibit or enhance, depending on the nature of the protein and the disease pathway, its function as a protein.”
As mentioned in last month’s article, through their Utility HPC platform, Cycle Computing aims to reduce computing time, resulting in lower costs for the clients.
Of course, while these tests are intriguing and the protein simulations are useful, what really showcases their worth is if such simulations result in the development of cancer drugs that otherwise would not have been possible. According to Stowe, Novartis proved that usefulness.
Two years ago, Novartis ran a 30,000-core Intel Xeon system setup on AWS via Cycle. They announced earlier this year that as a result of those computations, they found three compounds of interest from a drug target perspective.
These things often take time to verify as the drug trial process, which often involves clinical trials with cancerous patients, cannot yet be simulated via a supercomputing cluster. “Part of the enemy is time,” Stowe said. “We’ve run these workloads for significant clients in the past like Novartis and Schrodinger, but oftentimes you don’t know the impact of them until many years have passed.”
One of the advantages here is the ability to access a relatively large cluster for a short amount of time, thus accruing significantly less computing expense. For companies who look to run those 11 hour bursts, it would seem that they may need to prepare their applications ahead of time.
However, as evidenced by the other use cases Stowe referenced, namely CAD/CAM from an engineering perspective and genomics, there exist companies who make consistent use of a Cycle-AWS cluster over the course of three months.
“They had about 1200 of these 576-core jobs that they needed to run, each of which had its own 100-GB dataset,” Stowe said of a company who operated ten clusters concurrently to accomplish those 576-core jobs. Stowe estimated that such a physical system would take nine months to build, whereas that Cycle-facilitated process only ran three months. In another example, Cycle assisted in completing a million compute hours in one week for under $20,000 for a genomics company whose problem was shown below.
As evidenced by Stowe’s use cases and as noted by Gentzsch in the presentation, high performance applications run in the cloud serve a greater use to mid-sized institutions with neither the time nor the funds to purchase and implement an HPC cluster on their own. Those mid-sized institutions constitute a good portion of the 475 participants and 85 ‘teams’ that have participated in the HPC Experiment over its first three rounds.
Round four begins this month, and a more in-depth update will be provided within the next month in an article from Gentzsch. The full presentation from Stowe, Gentzsch, and Yenier can be found below.