Popular Post


Tuesday, September 25, 2012

Wavelets and Anomaly Detection - another CMG'12 paper

In the CMG'12 conference agenda I see another interesting paper appeared and it is written by Dima Seliverstov who I have met before in another CMG conference and mentioned in this blog

His paper is called 
"Introduction to Wavelets and their Application for Computer Performance Trend and
Anomaly Detection"

The very short abstract is 
"In this session, I will present a technique to identify trends and anomalies in Performance data using wavelets. I will answer the following questions: Why use Wavelets? What are Wavelets? How do I use them?"

CMG'12 conference Monday: 2:45 pm ‐ 3:45 pm Subject Area: Capacity Planning

I plan to attend. You?

Monday, September 17, 2012

LinkedIn Discussion around Trubin's Availability Formula

The previous post "Cluster Availability 9's Equation" triggered a very good discussion on LinkedIn Continuous Availability Forum.  It currently has 19 comments (!)... I plan to re-post some comments from the discussion here in my blog. (UPDATE: it is re-posted here)

BTW in one of the comments Bill Highleyman (co-author of the Breaking the Availability Barrier) pointed on the mistake in my formula which I corrected by replacing "n+n" with "mn". He also provided the excellent resource about availability calculation where he writes articles at the The Geek Corner for Availability Digest. One of the articles there extends the subject of this ( and couple previous) post and called: "Calculating Availability – Redundant Systems " 

As I suspected my formula ("Trubin law") is just a particular case of more generic rule Bill Highleyman formulates in that article. That says:

"... Adding a spare node adds the number of nines associated with that node to the system
availability but reduced by the increase in failure modes.

That is, adding an additional spare node adds the number of 9s of that node to the system
availability – almost. This improvement in availability is reduced a bit by the increase in the
number of failure modes in the system. More nodes mean more failure modes..."


Friday, September 14, 2012

Cluster Availability 9's Equation

Based on the "Trubin" Law (see my previous post) each additional node adds one more 9's to overall cluster availability. That is exactly true only if the single node has only one 9's (A=0.9), which the above "Trubin"'s equation shows.

But how that would work for other single node availability numbers? What if that has two or three 9's? I have generalized my previous equation to cover that and it shows that the cluster availability number of 9's will be increasing in arithmetic progression (sequence)!
Check more in the next post "LinkedIn Discussion around Trubin's Availability Formula"

Wednesday, September 12, 2012

The Right Number of Cluster Redundancy to Achieve the Availability Goal. Trubin's Law #4!

...Still playing with the availability algebra exercise, that I started in my other post:  How to Calculate Availability of Clustered Infrastructure for Multi-Tier Application
I have built the following two charts to see how the same availability goal can be achieved by different number of redundancy.

That of course possible if the less redundant configuration has more available individual components.

By the way, interesting that if the component availability has one 9's  (90.00%), then increasing redundancy on +1 gives the additional one 9's for the cluster availability. 

So, the equation
has the following solution:  A=0.9  for any integer n within the interval (0,)

It is probably the known law in the availability algebra. If not let's call that Trubin's law #4! ;)
This Saga of 9's continuous in the next post:

Cluster Availability 9's Equation

Friday, September 7, 2012

EV Based Trend Display - Trubin's Criterion!

Very typical task is to display trend direction on a dashboard type of report. Here is an example from SAS website:

But how correctly to choose the direction? Based on what? My suggestion is to use Exception Value (EV) meta metric (Check how that should be calculated here: EV-Control Chart). Indeed, for the given most recent time period (e.g. day, week or month) EV will be greater than 0 , if more UCL crossings happened than LCL crossings; EV  will be 0 if everything occurred within UCL-LCL band; and EV will be less than 0, if more LCL crossings happened than UCL crossings. 

That criterion have been already used to detect most recent trends, but the same way that could be used to choose the right direction for the trend arrow on dashboards!

So the Trubin's Criterion would be: