Intelligent management of virtualized resources for cloud database systems
View Patent ↗Systems and methods are disclosed to manage resources in a cloud-based computing system by generating a model of a relationship between cloud database resources and an expected profit based on cloud-server system parameters and service level agreements (SLAs) that indicates profits for different system performances, wherein the model comprises a two level optimization/control problem, wherein model receives system metrics, number of replicas, and arrival rate as the multiple input; and dynamically adjusting resource allocation among different customers based on current customer workload and the expected profit to maximize the expected profit for a cloud computing service provider.
1. A computer-implemented method to manage resources in a cloud-based computing system, comprising:
a. generating by a processor a model of a relationship between cloud database resources and a predicted profit based on cloud-server system parameters and service level agreements (SLAs) that indicates profits for alternate system performances, wherein the model comprises a two level optimization and control problem and wherein model receives system metrics, number of replicas, and arrival rate as inputs; and
b. dynamically adjusting by a processor a system resource allocation among alternate customers based on current customer workload and the predicted profit to maximize the expected profit for a cloud computing service provider while conforming to the SLA requirements.
2. The computer-implemented method of claim 1 , comprising predicting the profit under alternate system resource allocations.
3. The computer-implemented method of claim 1 , comprising building a predictive model to determine a potential profit gain under alternate system resource allocations.
4. The computer-implemented method of claim 1 , comprising obtaining an initial performance predictive model that predicts the predicted profit under a predetermined resource allocation.
5. The computer-implemented method of claim 1 , comprising updating the predictive model with feedbacks learned from a real-time system resource allocation and a corresponding cost.
6. The computer-implemented method of claim 1 , comprising generating a decision strategy that considers the predicted profit and an expense for making a change in the system resource allocation.
7. The computer-implemented method of claim 1 , comprising changing the system resource allocation in each allowed time-interval.
8. The computer-implemented method of claim 1 , comprising collecting new observations on performance under a new setting of system resource allocation.
9. The computer-implemented method of claim 1 , comprising:
determining with a processor a weighted SLA penalty cost function as:
P w ( q,i )= P ( q,i )× w ( i )
where within the k-th interval, an average SLA penalty cost for the i-th class of clients is a total SLA penalty cost over the total number of queries L(i,k) during the interval, where P represents a penalty cost function and w represents a weight function;
determining an average cost function AC as:
AC
(
i
,
k
)
=
1
L
(
i
,
k
)
∑
T
×
(
k
-
1
)
≤
q
start
<
T
×
k
P
(
q
,
i
)
;
determining a total weighted SLA penalty cost for the k-th interval and N classes of clients as:
SLA
(
k
)
=
∑
i
=
1
N
∑
T
×
(
k
-
1
)
≤
q
start
<
T
×
k
P
w
(
q
,
i
)
=
∑
i
=
1
N
AC
(
i
,
k
)
×
L
(
i
,
k
)
×
w
(
i
)
;
and
minimizing the total weighted SLA penalty cost under the constraints on resources.
10. The computer-implemented method of claim 1 , comprising:
a. predicting the profit under alternate system resource allocations;
b. building a predictive model to determine potential profit gain under alternate system resource allocations;
c. obtaining an initial performance predictive model that predicts the predicted profit under a resource allocation;
d. updating the predictive model with feedbacks learned from a real-time system resource allocation and corresponding cost;
e. generating a decision strategy that considers the expected profit and expense for making a change in the system resource allocation;
f. changing the system resource allocation in each allowed time-interval; and
g. collecting new observations on performance under a new setting of system resource allocation.
11. A system, comprising:
a. a model generator to build a model of relationships between allocated resources and a predicted profit based on cloud-server system parameters and service level agreements (SLAs) that indicates profits for alternate system performances; and
b. a dynamic control module coupled to the model generator, the dynamic control module adjusting resource allocations among alternate customers based on current customer workload and the predicted profit.
12. The system of claim 11 , wherein the system parameter comprises CPU share, memory share, database replicas.
13. The system of claim 11 , comprising means for predicting the profit under alternate system resource allocations.
14. The system of claim 11 , comprising means for building a predictive model to determine potential profit gain under alternate system resource allocations.
15. The system of claim 11 , comprising means for obtaining an initial performance predictive model that predicts the predicted profit under a resource allocation.
16. The system of claim 11 , comprising means for updating the predictive model with feedbacks learned from a real-time system resource allocation and corresponding cost.
17. The system of claim 11 , comprising means for generating a decision strategy that considers the predicted profit and expense for making a change in system resource allocation.
18. The system of claim 11 , comprising:
a processor for determining a weighted SLA penalty cost function as:
P w ( q i )= P ( q,i )× w ( i )
where within the k-th interval, an average SLA penalty cost for the i-th class of clients is a total SLA penalty cost over the total number of queries L(i,k) during the interval, where P represents a penalty cost function and w represents a weight function;
means for determining an average cost function AC as:
AC
(
i
,
k
)
=
1
L
(
i
,
k
)
∑
T
×
(
k
-
1
)
≤
q
start
<
T
×
k
P
(
q
,
i
)
;
means for determining a total weighted SLA penalty cost for the k-th interval and N classes of clients as:
SLA
(
k
)
=
∑
i
=
1
N
∑
T
×
(
k
-
1
)
≤
q
start
<
T
×
k
P
w
(
q
,
i
)
=
∑
i
=
1
N
AC
(
i
,
k
)
×
L
(
i
,
k
)
×
w
(
i
)
;
and
means for minimizing the total weighted SLA penalty cost under the constraints on resources.
19. A system, comprising:
a. a model generator to build a model of relationships between allocated resources and a predicted profit based on cloud-server system parameters and service level agreements (SLAs) that indicates profits for alternate system performances; and
b. a dynamic control module coupled to the model generator, the dynamic control module adjusting resource allocation among alternate customers based on current customer workload and the predicted profit
c. means for predicting the profit under alternate system resource allocations;
d. means for building a predictive model to determine potential profit gain under alternate system resource allocations;
e. means for obtaining an initial performance predictive model that predicts the predicted profit under a resource allocation;
f. means for updating the predictive model with feedbacks learned from a real-time system resource allocation and corresponding cost;
g. means for generating a decision strategy that considers the predicted profit and expense for making a change in the system resource allocation;
h. means for changing the system resource allocation in each allowed time-interval; and
i. means for collecting new observations on performance under a new setting of system resource allocation.