Telemetry-based method and system for predictive resource scaling
A telemetry-based method and system is provided for predictive resource scaling by mining useful data based on communication behaviour such as human calling patterns, call quality, and integration. In an embodiment, the telemetry-based system comprises at least one predictor for generating a first scaling prediction value based on a first target process goal and a second scaling prediction value based on a second target process goal; an auto-scaler for auto-scaling resources based on the first scaling prediction value and second scaling prediction value; and a regulator for monitoring the first scaling prediction value and second scaling prediction value and auto-scaling of resources by the auto-scaler and applying an objective function indicating total normalized system performance to a reinforcement learning algorithm for improving prediction accuracy of the at least one predictor.
1. A telemetry-based system for predictive resource scaling, comprising:
at least one predictor for generating a first scaling prediction value based on a first target process goal and a second scaling prediction value based on a second target process goal;
an auto-scaler for auto-scaling resources based on the first scaling prediction value and second scaling prediction value;
a regulator for monitoring the first scaling prediction value and second scaling prediction value and auto-scaling of resources by the auto-scaler and applying an objective function indicating total normalized system performance to a reinforcement learning algorithm for improving prediction accuracy of the at least one predictor, wherein the first target process goal is desired CPU utilization and the second target process goal is desired network utilization based on application domain related data;
a plurality of predictive auto-scaling groups configured to collect historical data and real-time data relating to the first target process goal and the second target process goal; and wherein said resources are virtual machine instances;
wherein at least one of the predictive auto-scaling groups comprises one of either an application serve, Session Initiation Protocol (SIP) server or Real-Time Transport Protocol (RTP) engine; and wherein the at least one predictor is configured to generate the second scaling prediction value by:
receiving historical data relating to application level performance,
formatting the historical data according to a data model of prediction,
training on the formatted historical data using the reinforcement learning algorithm; and is further configured to generate the first scaling prediction value by:
receiving the real-time data,
formatting the real-time data according to the data model of prediction,
training on the formatted historical data using the reinforcement learning algorithm, and in response sending a scaling signal to the auto-scaler for auto-scaling resources based on CPU utilization and application level performance.
2. The telemetry-based system of claim 1 , wherein the at least one predictor is further configured to receive statutory holiday calendar data and train on the statutory holiday calendar data in addition to formatted historical data.
3. The telemetry-based system of claim 1 , wherein the at least one predictor is further configured to receive regional major market index data and train on the regional major market index data in addition to formatted historical data.
4. The telemetry-based system of claim 1 , wherein the at least one predictor is further configured to receive social media data and train on the social media data in addition to formatted historical data.
5. The telemetry-based system of claim 1 , wherein the objective function indicating total normalized system performance is expressed as:
TperfCloud
=
(
∑
k
=
1
m
100
r_originate
(
k
)
)
/
nfs
+
(
∑
k
=
1
m
1
0
0
r_new
_dialog
(
k
)
)
/
nkm
+
(
∑
k
=
1
m
rtp_audio
_in
_mos
(
k
)
5
)
/
nrtp
,
where:
m is number of calls
nfs is number of virtual machine instances in the application server predictive auto-scaling group
nkm is number of virtual machine instances in the SIP server predictive auto-scaling group
nrtp is number of virtual machine instances in the RTP engine predictive auto-scaling group
r_originate(k) is response time to an originate command for a call k, as a measurement of a call K originated in the application server
r_new_dialog(k) is response time to a new dialog for the call k, as a measurement of a call K created in the SIP server
rtp_audio_in_mos(k) is a mos value of the call k, as a measurement of call K quality through the RTP engine, wherein a value range [0 . . . 5], 5 corresponds to best quality.
6. A telemetry-based system for predictive resource scaling, comprising at least one predictor for generating a first scaling prediction value based on a first target process goal and a second scaling prediction value based on a second target process goal;
an auto-scaler for auto-scaling resources based on the first scaling prediction value and second scaling prediction value; and
a regulator for monitoring the first scaling prediction value and second scaling prediction value and auto-scaling of resources by the auto-scaler and applying art objective function indicating total normalized system performance to a reinforcement learning algorithm for improving prediction accuracy of the at least one predictor,
wherein the first target process goal is desired CPU utilization and the second target process goal is desired network utilization based on application domain related data; and
wherein said resources are virtual machine instances, and
further comprising a plurality of predictive auto-scaling groups configured to monitor system hardware for autoscaling virtual machine instances based on traffic capabilities to achieve said desired CPU utilization.
7. The telemetry-based system of claim 6 , wherein the virtual machine instances include at least one of an application server, a Session Initiation Protocol (SIP) serves and a Real-Time Transport Protocol (RTP) engine.
8. The telemetry-based system of claim 7 , wherein the application server provides historical data relating to call legs.
9. The telemetry-based system of claim 8 , wherein the historical call data includes at least one of call quality, speed of answer and abandon rate.
10. The telemetry-based system of claim 7 , wherein the Session Initiation Protocol (SIP) server provides historical data relating to SIP messages.
11. The telemetry-based system of claim 10 , wherein the historical data relating to SIP messages includes at least one of request timeouts, gateway timeouts and busy.
12. The telemetry-based system of claim 7 , wherein the Real-Time Transport Protocol (RTP) engine provides historical data relating to at least one of call quality and error logs.
13. The telemetry-based system of claim 6 , further comprising a plurality of predictive auto-scaling groups configured to collect historical data and real-time data relating to the first target process goal and the second target process goal.
14. The telemetry-based system of claim 13 , wherein at least one of the predictive auto-scaling groups comprises one of either an application server, Session Initiation Protocol (SIP) server or Real-Time Transport Protocol (RTP) engine.