US 8,620,921 B1 · Modeler for predicting storage metrics

Abstract

Described herein is a system and method for dynamically managing service-level objectives (SLOs) for workloads of a cluster storage system. Proposed states/solutions of the cluster may be produced and evaluated to select one that achieves the SLOs for each workload. A planner engine may produce a state tree comprising nodes, each node representing a proposed state/solution. New nodes may be added to the state tree based on new solution types that are permitted, or nodes may be removed based on a received time constraint for executing a proposed solution or a client certification of a solution. The planner engine may call an evaluation engine to evaluate proposed states, the evaluation engine using an evaluation function that considers SLO, cost, and optimization goal characteristics to produce a single evaluation value for each proposed state. The planner engine may call a modeler engine that is trained using machine learning techniques.

Claims20

Dependent claims

A system for training a modeler for a storage system providing service to a plurality of workloads comprising a foreground workload and a first set of at least two background workloads, the system comprising:

computer hardware configured for:
receiving a plurality of data sets, each data set comprising values for a plurality of input metrics of the foreground workload, values for the plurality of input metrics of the first set of background workloads, and a value for an output metric of the foreground workload;
for each data set, summarizing the values received for each input metric of the first set of background workloads to produce a representative value for each input metric of the first set of background workloads;
producing a plurality of tuple sets, each tuple set representing a data set, each tuple set comprising values for the plurality of input metrics of the foreground workload, the representative values for the plurality of input metrics of the first set of background workloads, and a value for the output metric of the foreground workload;
and
processing the plurality of tuple sets using machine learning to produce a function that maps the plurality of input metrics of the foreground workload and the first set of background workloads to the output metric of the foreground workload.

3 more elements

A non-transitory computer readable medium having instructions stored thereon when executed by a processor, trains a modeler for a storage system providing service to a plurality of workloads comprising a foreground workload and a first set of at least two background workloads, the computer readable medium comprising sets of instructions for:

receiving a plurality of data sets, each data set comprising values for a plurality of input metrics of the foreground workload, values for the plurality of input metrics of the first set of background workloads, and a value for an output metric of the foreground workload;
for each data set, summarizing the values received for each input metric of the first set of background workloads to produce a representative value for each input metric of the first set of background workloads;
producing a plurality of tuple sets, each tuple set representing a data set, each tuple set comprising values for the plurality of input metrics of the foreground workload, the representative values for the plurality of input metrics of the first set of background workloads, and a value for the output metric of the foreground workload;
and
processing the plurality of tuple sets using machine learning to produce a function that maps the plurality of input metrics of the foreground workload and the first set of background workloads to the output metric of the foreground workload.

2 more elements

A system for training a modeler for a storage system providing service to a plurality of workloads comprising a foreground workload, a first set of at least two background workloads, and a second set of at least two background workloads, the system comprising:

computer hardware configured for:
receiving a plurality of data sets, each data set comprising values for a plurality of input metrics of the foreground workload, values for the plurality of input metrics of the first set of background workloads, values for the plurality of input metrics of the second set of background workloads, and a value for an output metric of the foreground workload;
for each data set, summarizing the values received for each input metric of the first set of background workloads to produce a representative value for each input metric of the first set of background workloads and summarizing the values received for each input metric of the second set of background workloads to produce a representative value for each input metric of the second set of background workloads;
producing a plurality of tuple sets, each tuple set representing a data set, each tuple set comprising values for the plurality of input metrics of the foreground workload, the representative values for the plurality of input metrics of the first set of background workloads, the representative values for the plurality of input metrics of the second set of background workloads, and a value for the output metric of the foreground workload;
and
processing the plurality of tuple sets using machine learning to produce a function that maps the plurality of input metrics of the foreground workload and the first and second sets of background workloads to the output metric of the foreground workload, the function for predicting a value of the output metric of the foreground workload based, at least in part, on values for the plurality of input metrics of the foreground workload and the first and second sets of background workloads.

3 more elements

A non-transitory computer readable medium having instructions stored thereon when executed by a processor, trains a modeler for a storage system providing service to a plurality of workloads comprising a foreground workload, a first set of at least two background workloads, and a second set of at least two background workloads, the computer readable medium comprising sets of instructions for:

receiving a plurality of data sets, each data set comprising values for a plurality of input metrics of the foreground workload, values for the plurality of input metrics of the first set of background workloads, values for the plurality of input metrics of the second set of background workloads, and a value for an output metric of the foreground workload;
for each data set, summarizing the values received for each input metric of the first set of background workloads to produce a representative value for each input metric of the first set of background workloads and summarizing the values received for each input metric of the second set of background workloads to produce a representative value for each input metric of the second set of background workloads;
producing a plurality of tuple sets, each tuple set representing a data set, each tuple set comprising values for the plurality of input metrics of the foreground workload, the representative values for the plurality of input metrics of the first set of background workloads, the representative values for the plurality of input metrics of the second set of background workloads, and a value for the output metric of the foreground workload;
and
processing the plurality of tuple sets using machine learning to produce a function that maps the plurality of input metrics of the foreground workload and the first and second sets of background workloads to the output metric of the foreground workload, the function for predicting a value of the output metric of the foreground workload based, at least in part, on values for the plurality of input metrics of the foreground workload and the first and second sets of background workloads.

2 more elements

Dependencies are inferred from the claim text.

Modeler for predicting storage metrics

Life of this patent

Modeler for predicting storage metrics

Life of this patent

Abstract

Claims20