Google has introduced a new compute service on the Google Cloud Platform that costs 70 percent less on average than an equivalent standard instance in the same configuration on the Google Compute Engine.
The catch with the Google Compute Engine Preemptible Virtual Machine? Google can shut down the job at any time.
There are a variety of computer tasks that fit nicely into this pricing model, Google Senior Product Manager Paul Nash said in a blog post.
The service, now in beta, would be good for fault-tolerant workloads that can be distributed easily across multiple virtual machines. Although jobs such as data analytics, genomics, and simulation and modeling can require lots of computational power, they can run periodically, or even if one or more nodes they're using goes offline.
Google's budget service is somewhat similar to Amazon Web Service's Spot Instances, also designed for jobs that can be interrupted. AWS' model is different because its price can fluctuate according to demand, whereas Google's prices are fixed. The Compute Engine Preemptible Virtual Machine can cost as little as US$0.01 per instance per hour.
To provide the service, Google is using leftover capacity in its data centers. If demand for its services spikes, the preemptible virtual machines (VMs) get bumped. Users are given a 30 second warning, which should give the application time to save its current state and work. No preemptible VM can run for more than 24 hours straight.
Some Google Cloud customers already plan to use the technology. Satellite imagery processing startup Descartes Labs recently required 30,000 processors to churn through a petabyte of NASA imagery. The lower costs may help the company work with even larger sets of data. Financial investment firm Citadel is another big user of cloud computing that will put these instances to work.
Preemptible VMs are started up and managed in the same manner as other Google Cloud Engine VMs, and existing tools and scripts can be used with them. Users of big data-analysis Hadoop software can also specify, through a single command, a given percentage of preemptible machines to use on their running workloads.