This week HPC cloud software specialist Cycle Computing announced that its full suite of products can now be used to spin up clusters on Google’s cloud platform. As testament to the new partnership, Cycle leveraged Google Compute Engine (GCE) to run a 50,000-core cancer gene analysis workload for the Broad Institute.
As Cycle Computing explains, Broad’s Cancer Group approached them with the need to perform a highly-complex genome analysis. The researchers already had powerful processing systems in-house, but running the analysis would take months and would require extensive coordination.
The decision was made to utilize the newly-launched “preemptible virtual machine” instances on GCE to further their cancer research. Preemptible VMs are Google’s answer to competitor Amazon’s spot instances. The preemptible instances are 60-70 percent cheaper than their on-demand counterparts. The catch is that Compute Engine can terminate (preempt) these instances at any time and there are a finite number available.
For applications that are “interruption friendly” (aka fault-tolerant), preemptible VMs offer a nice discount, and as Cycle explains, its software handles resiliency, enabling the orchestration of “clustered applications at any scale.”
Both classic “big compute” jobs as well as batch processing jobs can run on preemptible instances. If some instances terminate during processing, the job slows but does not completely stop.
Cycle expects the following applications will stand to benefit from preemptible VMs:
- Computational chemistry
- Needle-in-a-haystack simulations
- Financial pricing, back testing, modeling
- Genomics, bioinformatics, proteomics
- Insurance risk management
- Rendering, media encoding
- Hadoop, Spark, Redis, other IoT processing frameworks
Enabling greater access to utility-scale computing has always been the primary mission of Cycle Computing. The company has until now relied solely on Amazon’s cloud cycles, but by expanding its partner ecosystem it can better match and meet its customer needs. Recall that Broad and Google were already collaborating to develop new tools to facilitate and propel biomedical research. And in June, Broad Institute’s Genome Analysis Toolkit, or GATK, became available on Google Cloud Platform, as part of Google Genomics.
Cycle CEO Jason Stowe said Cycle doesn’t one recommend vendor over another, and that the applications cited are also well suited for AWS spot instances. “We provide tools that allow companies to benchmark their workloads on differing infrastructure and to be able to run them in production quality fashion; we stay out selection decisions. We obviously tell customers the options they have but we follow the customer.”
In general, he said, “Throughput-oriented stateless workloads tend to work well on that type of infrastructure and are definitely able to run on both Google GCE preemptible VMs and AWS spot instances.” The costs benefits can be substantial.
Broad’s Cancer Program has data sets pertaining to hundreds of cancer cell lines with information about genetic mutations, gene expression, and molecular interaction. Each level of data is massive in its own right, but exposing the hidden connections between these layers requires a comprehensive analysis. These relationships act as signposts directing the Cancer Program toward future research endeavors.
The scale of Broad’s scientific workload was not unfamiliar to Cycle, a company that prides itself on inspiring researchers to ask the “big questions” without regard to the limits of computing power. As Cycle describes it, this was a project that was at risk of not going forward if limited to available local resources.
“These types of analyses provide the clues that can lead to breakthroughs in disease research, such as cancer research, and this kind of cloud-based infrastructure helps us remove some of the local computing barriers that can stand in the way,” said Chris Dwan, acting director of information technology at the Broad Institute. “Flexible processing power allows us to think on a much larger scale.”
Revealing this map requires compute-intensive machine learning algorithms, the kind that would take months to execute on Broad’s on-premise system. The researchers already had the workload set up to run on an existing cloud-based StarCluster framework, so the challenge was to get this working on Google.
Cycle connected its CycleCloud to Google Cloud Platform, and ensured that its workload placement, data schedule, and at-scale computing capabilities were available on Google. Cycle says they were able to get this job up and running at moderate scales in 90 minutes using Cycle’s automation and orchestration tools as well as their cluster containers.
“The porting process for CycleCloud was very easy to accomplish. We were even able to simplify some of our existing code, because Google features like per-minute billing mean that we don’t have to worry about optimizing usage for hourly charges,” said Rob Futrick, chief technology officer for Cycle Computing.
Finding that the application hit its scaling sweet spot at about 50,000 cores, the Cycle team set the cluster to autoscale to 51,200 cores, requiring 3,210 16-core instances, using a mix of both n1-highmem and n1-standard types. Provisioned for less than the cost of a single server, this petascale cluster enabled Broad’s Cancer Group to complete their mapping workload in one afternoon. And as it so happens, some of the instances were preempted, but CycleCloud automatically reconfigured the cluster sans nodes, so the jobs continued.
After about six hours of computation, Broad’s map was complete. Analysis and curation will reveal the full extent of the relationships that were uncovered.