Automatic placement decisions for running incoming workloads on a datacenter infrastructure
A system and method place an incoming workload within a data center having infrastructure elements (IEs) for execution. Instrumentation data are collected for both individual IEs in the data center, and workload instances executing on each of these IEs. These data are used to train a future load model according to machine learning techniques, especially supervised learning. Future loads, in turn, are used to train a ranking model that ranks IEs according to suitability to execute additional workloads. After receiving an incoming workload, the first model is used to predict, for each IE, the load on its computing resources if the workload were executed on that IE. The resulting predicted loads are then fed into the second model to predict the best ranking of IEs, and the workload is placed on the highest-ranked IE that is available to execute the workload.
1. A computerized system for starting execution of a given workload in a data center having a plurality of infrastructure elements, the computerized system comprising:
a data receiver for receiving, from a workload manager, data indicating the given workload;
a future load predictor, for each infrastructure element in the plurality of infrastructure elements, using a first model trained using machine learning to predict a load, during a future time window for each of a plurality of computing resources, that would occur if the given workload were executed using that infrastructure element;
a placement ranking predictor, using a second model trained using machine learning to predict, as a function of the predicted future loads, a ranking of infrastructure elements that would be most suited to execute the given workload;
an infrastructure element selector for selecting one or more infrastructure elements according to the ranking; and
a data transmitter for transmitting, to the workload manager, data indicating the selected one or more infrastructure elements;
wherein the workload manager responsively starts execution of the given workload on the indicated one or more infrastructure elements.
2. The system of claim 1 , wherein either or both of the first model and the second model were trained, using supervised machine learning, on telemetry data collected from the plurality of infrastructure elements over a plurality of time windows.
3. The system of claim 1 , wherein the computing resources include any combination of a required CPU time, a required memory space, and a required disk space.
4. The system of claim 1 , wherein using the first model comprises providing, to the first model, inputs comprising statistical measures of computing resources used during a current time window by (a) other instances of the given workload currently executing in the data center, and (b) each infrastructure element in the plurality of infrastructure elements.
5. The system of claim 1 , wherein using the first model comprises, when no instances of the given workload are currently executing in the data center, predicting a future load for each of the plurality of computing resources that corresponds to a historical average load.
6. The system of claim 1 , wherein selecting the one or more infrastructure elements according to the ranking comprises selecting a highest ranked infrastructure element that satisfies one or more workload acceptance criteria.
7. A method of starting execution of a given workload in a data center having a plurality of infrastructure elements, the method comprising:
for each infrastructure element in the plurality of infrastructure elements, using a first model trained using machine learning to predict a load, during a future time window for each of a plurality of computing resources, that would occur if the given workload were executed using that infrastructure element;
using a second model trained using machine learning to predict, as a function of the predicted future loads, a ranking of infrastructure elements that would be most suited to execute the given workload;
selecting one or more infrastructure elements according to the ranking; and
starting execution of the given workload on the selected one or more infrastructure elements.
8. The method of claim 7 , wherein either or both of the first model and the second model were trained, using supervised machine learning, on telemetry data collected from the plurality of infrastructure elements over a plurality of time windows.
9. The method of claim 7 , wherein the computing resources include any combination of a required CPU time, a required memory space, and a required disk space.
10. The method of claim 7 , wherein using the first model comprises providing, to the first model, inputs comprising statistical measures of computing resources used during a current time window by (a) other instances of the given workload currently executing in the data center, and (b) each infrastructure element in the plurality of infrastructure elements.
11. The method of claim 7 , wherein using the first model comprises, when no instances of the given workload are currently executing in the data center, predicting a future load for each of the plurality of computing resources that corresponds to a historical average load.
12. The method of claim 7 , wherein selecting the one or more infrastructure elements according to the ranking comprises selecting a highest ranked infrastructure element that satisfies one or more workload acceptance criteria.
13. The method of claim 7 , wherein starting execution of the given workload comprises starting a new service on a computer server, or starting a new process in an existing virtual machine, or starting execution of a containerized process.
14. A non-transitory, computer-readable storage medium, in which is stored computer program code for performing a method of starting execution of a given workload in a data center having a plurality of infrastructure elements, the method comprising:
for each infrastructure element in the plurality of infrastructure elements, using a first model trained using machine learning to predict a load, during a future time window for each of a plurality of computing resources, that would occur if the given workload were executed using that infrastructure element;
using a second model trained using machine learning to predict, as a function of the predicted future loads, a ranking of infrastructure elements that would be most suited to execute the given workload;
selecting one or more infrastructure elements according to the ranking; and
starting execution of the given workload on the selected one or more infrastructure elements.
15. The storage medium of claim 14 , further comprising computer program code for training either or both of the first model and the second model using supervised machine learning on telemetry data collected from the plurality of infrastructure elements over a plurality of time windows.
16. The storage medium of claim 14 , wherein the computing resources include any combination of a required CPU time, a required memory space, and a required disk space.
17. The storage medium of claim 14 , wherein the computer program code for using the first model comprises computer program code for providing, to the first model, inputs comprising statistical measures of computing resources used during a current time window by (a) other instances of the given workload currently executing in the data center, and (b) each infrastructure element in the plurality of infrastructure elements.
18. The storage medium of claim 14 , wherein the computer program code for using the first model comprises computer program code for, when no instances of the given workload are currently executing in the data center, predicting a future load for each of the plurality of computing resources that corresponds to a historical average load.
19. The storage medium of claim 14 , wherein the computer program code for selecting the one or more infrastructure elements according to the ranking comprises computer program code for selecting a highest ranked infrastructure element that satisfies one or more workload acceptance criteria.
20. The storage medium of claim 14 , wherein the computer program code for starting execution of the given workload comprises computer program code for starting a new service on a computer server, or starting a new process in an existing virtual machine, or starting execution of a containerized process.