Granted Patent B2

US 12,608,233 · App. 18/016,982 · Granted Apr 21, 2026

Inventors: Gabriel Martin (Quintana de la Sirena, ES); Max Alt (San Francisco, CA)

Assignee: Advanced Micro Devices, Inc.

G06F9/5038G06F9/5072G06F2209/5019

Abstract

Claims (48)

1 . A method for recommending computing resources in a distributed computing system, the method comprising:

(a) gathering performance data by running a plurality of training computing jobs on a plurality of different computing devices in the distributed computing system;

(b) storing the performance data in a database;

(c) performing feature extraction on the gathered performance data to identify job features of the training jobs and system features of the computing devices;

(d) creating a matrix of the job features and the system features provided with the extracted performance data;

(e) generate performance estimates using a plurality of recommendation algorithms;

(f) cross validating the performance estimates to identify best estimates; and

(g) recommend computing resources for a next computing job using the best estimates.

2 . The method of claim 1 , further comprising:

executing a next computing job on the recommended computing resources; and

repeating steps (a)-(f) using new performance data generated from the execution of the next computing job.

3 . The method of claim 1 , further comprising aggregating the performance data in the database by job, app, queue, and app-queue.

4 . The method of claim 1 , wherein the performance data comprises instruction per second metrics.

5 . The method of claim 1 , wherein one of the features is cloud versus bare metal.

6 . The method of claim 1 , further comprising using similar queues to provide data for queues where insufficient performance data exists in the database to perform steps (c)-(f).

7 . The method of claim 1 , wherein at least one of the plurality of recommendation algorithms is K-nearest neighbors (KNN).

8 . The method of claim 1 , wherein the plurality of recommendation algorithms comprises content-based algorithms and collaborative filtering using different feature sets and different performance metrics.

9 . The method of claim 1 , wherein the performance data comprises job parameters comprising measures of CPU boundness, GPU boundness, MPI boundness, memory boundness, and network boundness.

10 . The method of claim 9 , wherein the performance data further comprises latency and bandwidth.

11 . The method of claim 1 , wherein the job features comprise libraries used, datasets used, application domain, and application metadata.

12 . The method of claim 1 , wherein the system features comprise a number of cores, a core type, an available memory amount, and a maximum core clock speed.

13 . The method of claim 1 , wherein the system features comprise storage type and network type.

14 . The method of claim 1 , wherein the system features comprise IO throughput.

15 . The method of claim 1 , wherein the system features comprise memory bandwidth and latency.

16 . The method of claim 1 , wherein the system features comprise benchmark scores.

17 . A system having at least one processor, the system configured to:

gather performance data by running a plurality of training computing jobs on a plurality of different computing devices in a distributed computing system;

store the performance data in a database;

perform feature extraction on the gathered performance data to identify job features of the training jobs and system features of the computing devices;

create a matrix of the job features and the system features provided with the extracted performance data;

generate performance estimates based on a plurality of recommendation algorithms;

cross-validate the performance estimates to identify best estimates; and

recommend computing resources for a next computing job using the best estimates.

18 . A system having at least one processor, the system configured to:

analyze gathered performance data to create aggregated performance values, wherein the gathered performance data is obtained by running a plurality of test jobs on a plurality of different subsets of computing resources;

perform feature extraction on the aggregated performance values to identify job features of the test jobs and system features of the computing resources;

populate a matrix with the job features and the system features based on the aggregated performance values;

create a recommender using the matrix;

characterize a new job to determine which of the test jobs is most similar based on resource utilization; and

select via the recommender an optimal subset of the computing resources to execute the new job based on the aggregated performance values for the most similar test job.

19 . The system of claim 18 , further configured to:

execute the new job;

gather new performance data for the new job;

create new aggregated performance values based on the gathered new performance data;

update the matrix with the new aggregated performance values; and

update the recommender using the updated matrix.

20 . The system of claim 18 , further configured to select via the recommender a subset of system resources that comprises more of a specific type of computing resource in response to the new job showing intensive utilization of the specific type of computing resource.

21 . The system of claim 18 , wherein if the new job's performance on a first computing system characterizes similarly to one of the plurality of test job's performance on the first computing system, then the recommender uses the test job's performance on a second computing system to estimate a performance of the new job on the second computing system.

Assignments (2)

Continuity (2)

References Cited (12)