Adaptable response time prediction for storage systems under variable workloads
One example method includes running a workload through a trained open-set classification model, recovering, as a result of the running, a class and an open-setness score corresponding to the workload, determining, based on the class and the open-setness score, whether the workload is new, and when the workload is determined to be new, starting a new cluster that includes the workload. A response time predictor model may be used to predict a response time associated with the new workload.
1 . A method, comprising:
running a workload comprising telemetry data including workload characteristics and system configuration parameters of a data storage system through a trained open-set classification model that has been trained using clustered workloads and supervised with cluster identifiers;
recovering, as a result of the running, a class and an open-setness score corresponding to the workload;
calculating a membership score for each of a plurality of known clusters based on a distance between the workload and a centroid of each known cluster;
aggregating the open-setness score with membership scores to determine whether the workload corresponds to a previously unseen workload type;
determining, based on the aggregated open-setness score and the membership scores, whether the workload is new; and
when the workload is determined to be new, starting a new cluster that includes the workload and storing the workload in association with the new cluster for subsequent training of a response-time predictor model.
2 . The method as recited in claim 1 , wherein the workload comprises IO operations associated with a data storage system.
3 . The method as recited in claim 1 , wherein after training based on stored workloads associated with the new cluster, a response time predictor model associated with the new cluster performs an inferencing process that determines a response time, of a system, associated with the workload.
4 . The method as recited in claim 1 , wherein when the workload is determined not to be new, the workload is assigned to a best matching cluster.
5 . The method as recited in claim 1 , wherein the workload comprises workload characteristics and a system configuration for a system that processes the workload.
6 . The method as recited in claim 1 , wherein when the workload is determined not to be new, and no response time predictor model exists for a best matching cluster corresponding to the workload, an output of ‘no possible prediction’ is generated when a number of workloads in the best matching cluster is below a threshold.
7 . The method as recited in claim 1 , wherein when the workload is determined not to be new, and no response time predictor model exists for a best matching cluster corresponding to the workload, a response time predictor model is trained for the workload when a number of workloads in the best matching cluster is above a threshold.
8 . The method as recited in claim 1 , wherein the new cluster is specifically associated with a response time predictor model after the response time predictor modal is trained using workloads stored in association with the new cluster operable to perform an inferencing process that determines a response time, of a system, associated with the workload.
9 . The method as recited in claim 1 , wherein the open-setness score reflects a relative newness of the workload.
10 . The method as recited in claim 1 , wherein the open-set classification model is operable to classify workloads into clusters.
11 . A non-transitory storage medium having stored therein instructions that are executable by one or more hardware processors to perform operations comprising:
running a workload comprising telemetry data including workload characteristics and system configuration parameters of a data storage system through a trained open-set classification model that has been trained using clustered workloads and supervised with cluster identifiers;
recovering, as a result of the running, a class and an open-setness score corresponding to the workload;
calculating a membership score for each of a plurality of known clusters based on a distance between the workload and a centroid of each known cluster;
aggregating the open-setness score with the membership scores to determine whether the workload corresponds to a previously unseen workload type;
determining, based on the aggregated open-setness score and the membership scores, whether the workload is new; and
when the workload is determined to be new, starting a new cluster that includes the workload and storing the workload in association with the new cluster for subsequent training of a response-time predictor model.
12 . The non-transitory storage medium as recited in claim 11 , wherein the workload comprises IO operations associated with a data storage system.
13 . The non-transitory storage medium as recited in claim 11 , wherein after training based on stored workloads associated with the new cluster, a response time predictor model associated with the new cluster performs an inferencing process that determines a response time, of a system, associated with the workload.
14 . The non-transitory storage medium as recited in claim 11 , wherein when the workload is determined not to be new, the workload is assigned to a best matching cluster.
15 . The non-transitory storage medium as recited in claim 11 , wherein the workload comprises workload characteristics and a system configuration for a system that processes the workload.
16 . The non-transitory storage medium as recited in claim 11 , wherein when the workload is determined not to be new, and no response time predictor model exists for a best matching cluster corresponding to the workload, an output of ‘no possible prediction’ is generated when a number of workloads in the best matching cluster is below a threshold.
17 . The non-transitory storage medium as recited in claim 11 , wherein when the workload is determined not to be new, and no response time predictor model exists for a best matching cluster corresponding to the workload, a response time predictor model is trained for the workload when a number of workloads in the best matching cluster is above a threshold.
18 . The non-transitory storage medium as recited in claim 11 , wherein the new cluster is specifically associated with a response time predictor model after the response time predictor model is trained using workloads stored in association with the new cluster operable to perform an inferencing process that determines a response time, of a system, associated with the workload.
19 . The non-transitory storage medium as recited in claim 11 , wherein the open-setness score reflects a relative newness of the workload.
20 . The non-transitory storage medium as recited in claim 11 , wherein the open-set classification model is operable to classify workloads into clusters.