ANALYSIS FOR
ALL NODES

Alarm triggering condition: The standard deviation of the forecast
has to be smaller than the residual (F-y) in 70% of the points in the last 2,5
hours.

Click on the name to get the excel file.

No alerts!!

NO
ALARMS!!! PERFECT!!

NO
ALARMS!!!!!

NO
ALARMS!!!

2 ALARMS!!! It’s one of them false??? Maybe not.

NO
ALARMS

NO
ALARMS!!!

Total Failure!!!!!!! But it’s obvious since the first week
give the tendency. Not even a very little alpha and a high beta solve this.
This node is very particular. The values of dbcap finally tend to 368.3 and stay there for several
weeks without any change. If we start our analysis from the 5^{th} week
we won’t get any alarm

ONE ALARM!!!!! PERFECT!!!

ONE ALARM IN THE

I am not completely sure if this should be taken as 1 or many ALARMS. Let’s say that is only 1. Because I am always assuming that if the bad condition repeats itself during consecutive days, it should be part of the same error.

NO ALARMS!!!

The same thing than SOX, and it fails pretty badly.

NO ALARMS!!!!!!!!!!!!!!!!!!!!!!!!

** MCS**… This
one needs a little more attention

I am pretty sure that this is a good alarm. It seems like it is from my point of view.

Now THIS??? It is clearly a change in the normal weekly behavior, but is it really an Alarm? I think it is. Maybe not….

And this!!!! We can see 1 clear step change at the beginning, and then many alarms which are all part of a big failure… And at the end.. ok let’s call it a FALSE alarm.

NO ALARMS!!!

** LSA**…
closer look here

We get this… I think it’s OK

I don’t know why, but we don’t get the big one… 1 MISSED

3 more alarms… the second one is OK, I think the third one too. The first one, I don’t know.

4 alerts…

And later..

We get a double alert at the end.

Let’s look at this closer: are these good alerts?

If we go to the smallest one. On 23/8/2004:

This bad performance last more than 12 hours, so maybe it is a good alert.

__CCSORNL__

NO ALARMS!!!!

NO ALARMS!!!!

The same problem than SOX…

NO ALARMS”!!!!!

NO ALARMS HERE….

BUT LATER ON…

ONE MISSED…

NO ALARMS!!!!!!

2 Alarms… 1 false… the other, seems to be false too…

NO ALARMS!!!

NO ALARMS!!! PERFECT

NO ALARMS!!! Good…

NO ALARMS!!!!

We get many diurnal alerts that are impossible for HW to predict. These diurnal effects don’t happen in previous weeks.

Comments

We get almost no false alarms, except when the first weeks show really better performance in average than the weeks that follow or when the standard deviation is very small (constant dbcap many days). SO, when the data find a stable value, the standard deviation is so little, that any residual can cause an alarm.

Getting the real minimum MSE instead of a estimative value of this minimization could solve all these errors.

We have to remember that the Holt-Winters algorithm assumes that there will be a seasonal trend, and in the case of Cesnet or SOX, there is not a seasonal trend since the values of the dbcap tend to be constant on the long term and very different compared with the first couple of weeks.

We generally don’t miss the important step changes.

Without counting the particular SOX situation we get:

32 alarms in 3 months on 29 nodes

15 good

6 doubtful…

3 missed

2 false…

6 Diurnal on SDSC