Five Easy Steps to Understanding The Impact of Using
Reactive Analytics for Managing IT Infrastructure
WATCH THE VIDEO
Take 3 minutes to understand why you can't control and
optimize infrastructure with reactive tools
UNDERSTAND THE SCIENCE OF CONTROL THEORY
Controlling IT is a lot like controlling robots…
Standard robotic control theory makes a distinction between reactive control, which is fast but simple, and predictive control, which is much smarter. One is designed to react quickly to keep you out of trouble, while the other is designed to understand exactly what you are trying to accomplish, and provide precise actions to get you there.
From “Behavior-Based Robotics” by R. Arkin, MIT Press, 1998
Predictive (Deliberative)Designed to meet your goals by having a
deep understanding of data center supply
and demand and detailed policies that drive
ReactiveDesigned to recover from problems by reacting to events, using short windows of utilization data, but without an understanding of how to actually optimize the environment.
LISTEN TO THE EXPERTS
It's time to cut through the marketing hype.
Raouf Boutaba, Computer Science Professor at the renowned University of Waterloo and IEEE Fellow, is an expert in VM placement and resource optimization in virtual and cloud infrastructure. He recently evaluated and tested another product after it was discovered to be causing significant performance issues in a client environment.
Download his paper, to understand the causes of:
- Frequent VM placement oscillation
- Frequent VM migration
- Sub-optimal VM density
- A desired state which is not the optimal state of the environment
Read this research paper now.
Raouf Boutaba, Computer Science Professor
University of Waterloo & IEEE Fellow
ASK THE RIGHT QUESTIONS
Ask the right questions when using or evaluating a Reactive Analytics Tool.
Q.How are Reactive Analytic Tools different from VMware DRS + vROps 6.1? If they do the same thing, then why would I not go with the solution from the existing hypervisor vendor?
Q.How are business cycles accounted for, and how can servers be removed if periodic loads may require more/all of the existing server capacity? And how does it account for inbound applications that will require resources in the future?
Q.If I want a Reactive Analytic Tool to balance the CPU usage, memory, overcommit ratios, I/O levels, etc. across a set of physical servers, all at the same time, explain how exactly it will do this? How can a single utilization index properly represent all of these factors?
Q.What exists in the application to reduce software licensing costs beyond attempting to reduce the overall server footprint? Can it target a specific subset of licensed VMs and optimize them separately in order to give the true minimum footprint?
Q.Can it support future service requests and reserve capacity for inbound applications? Will it reject subsequent requests if outstanding bookings have a claim on the required resources?
TRY BEFORE YOU BUY
Verify the Research Findings
for Yourself. Try these simple
test cases to see the limitations.
1. The “Deadlock” Test
Set up a cluster with 2 hosts and create
4 VMs on each. Run load generators in the
VMs to create utilization levels that looks
Observe how nothing happens, even though a simple VM move will rebalance the systems. This is because the Utilization Index is based on only the busiest resource for each server, over-simplifying the problem and making it blind to what is actually happening.
2. The “Keep Your Job” Test
Create a VM with 2 virtual CPUs. Use a load generator to simulate a critical start-of-day business load by making the VM 100% busy from 6-8am each day. Be sure it has no load at other times.
Enable the tool and observe what it recommends (Note: this may take some time). Because it is using a simplistic percentile approach, it will think that the VM is over-sized (90th percentile utilization = 0%) and will recommend down-sizing the VM to 1 vCPU. This would be catastrophic, and will cause the workload to miss start of day. Business groups don’t like this.