BETA
This is a BETA experience. You may opt-out by clicking here

More From Forbes

Edit Story

How Advanced Modeling Of Data Centers Can Save Millions

This article is more than 10 years old.

VMware has made a fortune based on a simple premise. If you take the software running on 100 underutilized physical computers and put them in 100 virtual machines on 10 physical computers, you save a ton of money.

Of course you gain lots of other stuff from virtualization -- flexibility, speed of deployment, better reporting, etc. – but that massive reduction in hardware is often the most impressive accomplishment. What is less often considered in this equation is: Why was so much hardware deployed in the first place? The most common causes are ignorance and fear. People put applications on too much hardware because they don’t understand the operational characteristics of the workloads or the hardware and want to leave some headroom. Of course, later on, when the application is understood, the headroom isn’t removed, leaving a layer of fat in the data center.

But let’s think about this more deeply. Do you think the “headroom as insurance mentality,” a fear-based concept, is just limited to processing power? Perhaps the network devices, the storage networks, the firewalls, and so on all are over-provisioned? How would we know? How would we know what an adequate level of hardware and software is for any given task?

These last two questions are the core focus of a company called CloudPhysics, a Silicon Valley firm with an A-list set of investors (Kleiner Perkins, Mayfield) and a nerdy, super technical CEO named John Blumenthal.

In the abstract, CloudPhysics is one of a new breed of company like Opera Solutions, Palantir, RocketFuel, and others who are creating high-value offerings by combining into one product or service advanced data science, state of the art software development, and deep vertical expertise.

CloudPhysics has aimed its brainpower at data centers, initially those running VMware. If the company hits a home run, there will be massive implications for how data centers are planned and run.

The opportunity for CloudPhysics arose because more data is available. When computing becomes virtual, a wealth of data becomes available about how everything is running. “The hypervisor has introduced a normalized layer of data for the data center,” said Blumenthal. “You have one layer through which all of your resource consumption can be seen without having to go to each server or each vendor. Once that arrived, new forms of analysis became possible.”

CloudPhysics uses this data to create a model of the systems in the data center. Instead of attempting to boil the ocean and create a comprehensive model of all aspects of a data center, something that would not be easy or even possible in the short term, CloudPhysics models are targeted at answering questions that have high economic value. The models are focused on the behavior of specific systems. They are developed using statistical indicators that describe the system at large. In some cases, the behavior and resource consumption of important component parts such as SSD drives are modeled.

Here are a few of the questions that the models CloudPhysics has developed can answer:

  • What would be the effect of adding a server-side caching system?
  • Will my virtual machines successfully fail over if I change my high availability policy or provision new virtual machines in my high availability group?
  • Is it more cost effective to run my workloads in my data center or in one of the public cloud options like VMware’s vCHS or Amazon’s EC2?
  • What performance changes occurred within all the layers of my cluster at the time of a particular event, such as a configuration change? What performance changes will occur if a particular configuration change occurs?

The idea is that you can look at how your data center is running like a Formula 1 team looks at thousands of real-time metrics for a race car. With this data in hand, it becomes possible to make data-based decisions about how to design, deploy, and optimize all levels of the equipment and software instead of relying on estimates based on fear. The “headroom as insurance” mentality can be confidently discarded.

“The only way to do more with less today – and do it safely – is to improve the quality of decisions and processes in every IT operational use case,” said Blumenthal “By applying data science and simulation to IT Ops, CloudPhysics raises operations to radically new levels of efficiency and safety.”

CloudPhysics, a Software as a Service offering, harvests data in the following manner:

  • Clients deploy a lightweight CloudPhysics “observer” in their data center. The observer scans their virtualized environment continuously and sends data to CloudPhysics, where performance, configuration, and other data gathered is used to expose potential operational hazards and opportunities for cost savings.
  • Anonymized data from user environments is aggregated on the CloudPhysics platform, creating a global data set for modeling, simulations, and what-if scenarios. This data set can be used to create and improve the quality of statistical models.

CloudPhysics realizes that its capabilities and data will need to be used in many different ways, so the company has created an extensible architecture that allows its customers to create apps on top of the core platform to do specific tasks. The apps are called “cards” and are discoverable in the card store. At this point, users create such apps using a card builder, but it is likely in the future that the underlying APIs of the platform will be exposed to developers.

Blumenthal’s team can make this work because they come from three rarely combined areas: virtualization resource management, modeling and simulation, and data science.

But the implications of CloudPhysics go beyond optimizing the design of a data center. The detailed metrics tracked by the models mean that it will also be possible to very precisely monitor the normal ranges of every component. Abnormal behavior can be flagged. This type of analysis will be hugely important for detecting attacks through Advanced Persistent Threats and other means.

If CloudPhysics’s approach drives savings, it is a short trip to attacking other similar data-rich environments that need illumination.

“We continue to expand our modeling and simulation to new parts of cloud infrastructure with the help of the community and partners. While much of our current focus is on storage in VMware environments, we are currently exploring Hyper-V and OpenStack data collections,” said Blumenthal.

CloudPhysics represents the productization of advanced data science applied to high-value problems. It is an early example of what will likely be a wave of such companies. CloudPhysics represents a pattern for these new offerings in the way that data is used to construct just enough of a model to answer crucial questions. In addition, CloudPhysics shows how data can attack costly decisions based on fear, something that seems to have the potential for massive return in many different scenarios.

Follow Dan Woods on Twitter:

Dan Woods is CTO and editor of CITO Research, a publication that seeks to advance the craft of technology leadership. For more stories like this one visit www.CITOResearch.com.