Using variance to analyze DBMS performance

Fig.1. Short moving DBMS response time (median smoothing with a period of 10 minutes)

Basic statistical indicators of DBMS performance

Median smoothing, with a period of 1 hour.

Fig.3. Long rolling DBMS performance metrics

Fig.4. Long moving - sessions in a waiting state — Fig.4. Long moving – sessions in a waiting state

Fig.5. Correlation between performance and sessions in the standby state

Additional statistical indicators for analyzing DBMS performance

Performance Variance

Fig.6. Dispersion of long moving average

For each point t the sample variance is calculated for the input values (sample deviation squared) for the period [ t ; t-60 минут].

Fig.7. Tangent of the tangent to the DBMS performance dispersion point

Important clarifications on schedules

1. On the performance dispersion graph, there is clearly an inflection point in the graph.

2.Point in time t (60, marked on the graphs with a large yellow dot), in which a break in the dispersion graph appeared (Fig.6) and the break point of the graph of the tangent of the inclined tangent angle (Fig.7) occurred much earlier than a noticeable increase in response time (Fig.1, Fig.2 )

More detailed fragments of graphs

Fig.8. Break point of the DBMS performance dispersion graph

Fig.9. Tangent of the tangent to the performance dispersion graph

Fig. 10. Short sliding response time. The yellow dot is a kink in the performance dispersion graph. The red dot is an irreversible increase in response time.

Fig. 11. Long sliding response time. The yellow dot is a kink in the performance dispersion graph. The red dot is an irreversible increase in response time.

Results, conclusions and plans for the future

The condition of a break in the dispersion graph of the moving DBMS performance value is a sufficient condition for the degradation of DBMS performance.
The break in the dispersion graph of the moving DBMS performance value occurs before an irreversible increase in the DBMS response time during load testing.
To analyze the possibility of using the dispersion graph of the moving average of DBMS performance, additional observations are required under productive load conditions.
Interpretation and deeper analysis of the DBMS performance dispersion graph requires additional experimentation and time.
Perhaps in the course of experiments, it will be possible to obtain new metrics and increase the efficiency of statistical analysis of DBMS performance.