Granted Patent B2

US 12,579,070 · App. 18/527,295 · Granted Mar 17, 2026

Inventors: Amir Rosen (Haifa, IL); Shie Mannor (Haifa, IL); Sagi Lahav (Kiryat Bialik, IL); Gil Levy (Hod Hasharon, IL); Ariel Szapiro (Kfar Netter, IL)

Assignee: Mellanox Technologies, Ltd

G06F12/0862G06N20/00

Abstract

Claims (77)

1 . A method, comprising:

receiving data of a set of configurations of preprocessor engines;

receiving measurements of performance of a device executing benchmark applications while changing a configuration of preprocessor engines selected from the set of configurations of preprocessor engines;

defining an order of at least some of the configurations based on the measurements; and

providing a pruned set of configurations based on the defined order of the at least some configurations.

2 . The method according to claim 1 , further comprising:

reducing a collection of configurations of the preprocessor engines to yield the set of configurations of the preprocessor engines using a first configuration reduction method; and

reducing the set of configurations of the preprocessor engines to yield the pruned set of configurations using a second configuration reduction method different to the first configuration reduction method.

3 . The method according to claim 1 , wherein the preprocessor engines are prefetcher engines, and the configurations are different prefetcher engine configurations based on aggressiveness levels to be applied to the prefetcher engines.

4 . The method according to claim 3 , further comprising:

executing the benchmark applications while changing the aggressiveness levels of the prefetcher engines; and

measuring the performance of the device during the execution of the benchmark applications.

5 . The method according to claim 1 , wherein defining the order includes finding:

a best configuration of the configurations based on the measurements;

a second-best configuration of the configurations based on the best configuration being found to be the best configuration and based on the measurements; and

a third-best configuration of the configurations based on the best configuration and the second-best configuration being found to the best configuration and the second-best configuration, respectively, and based on the measurements; and

an N th best configuration of the configurations based on the best to N−1 best configurations of the configurations being found to be the best to N−1 best configurations, respectively, and based on the measurements.

6 . The method according to claim 5 , further comprising computing first averages of the vectors that indicate the performance of the device executing benchmark applications for respective ones of the configurations, wherein the finding includes: finding a first vector of the vectors having a highest average of the first averages; and defining the best configuration as one of the configurations corresponding to the first vector.

7 . The method according to claim 6 , wherein the finding the second-best configuration includes:

(a) selecting a first candidate vector of the vectors;

(b) performing elementwise maximum operations with the first candidate vector and the first vector yielding a first resultant vector;

(c) computing an average of the first resultant vector;

(d) repeating steps (a) to (c) for other first candidate vectors of the vectors yielding multiple first resultant vectors with multiple corresponding second averages; and

(e) finding the second-best configuration corresponding to a second vector of the first candidate vectors yielding a given first resultant vectors of the first resultant vectors having a highest average of the second averages.

8 . The method according to claim 7 , wherein the finding the third-best configuration includes:

(a) selecting a second candidate vector of the vectors;

(b) performing elementwise maximum operations with the second candidate vector and the given first resultant vector, or with the first vector and the second vector, yielding a second resultant vector;

(c) computing an average of the second resultant vector;

(d) repeating steps (a) to (c) for other second candidate vectors of the vectors yielding multiple second resultant vectors with multiple corresponding third averages; and

(e) finding the third-best configuration corresponding to one of the second candidate vectors yielding one of the second resultant vectors having a highest average of the third averages.

9 . The method according to claim 6 , wherein the finding the N th best configuration includes:

(a) selecting a candidate vector of the vectors;

(b) performing elementwise maximum operations with the candidate vector and at least one vector indicative of the configurations from the best configuration to an (N−1) th best configuration, yielding a resultant vector;

(c) computing an average of the resultant vector;

(d) repeating steps (a) to (c) for other candidate vectors of the vectors yielding multiple resultant vectors with multiple corresponding averages; and

(e) finding the N th best configuration corresponding to one of the candidate vectors yielding one of the resultant vectors having a highest average of the multiple corresponding averages.

10 . A system, comprising:

a processor to:

receive data of a set of configurations of preprocessor engines;

receive measurements of performance of a device executing benchmark applications while changing a configuration of preprocessor engines selected from the set of configurations of preprocessor engines;

define an order of at least some of the configurations based on the measurements; and

provide a pruned set of configurations based on the defined order of the at least some configurations; and

a memory to store data used by the processor.

11 . The system according to claim 10 , wherein the processor is to:

reduce a collection of configurations of the preprocessor engines to yield the set of configurations of the preprocessor engines using a first configuration reduction method; and

reduce the set of configurations of the preprocessor engines to yield the pruned set of configurations using a second configuration reduction method different to the first configuration reduction method.

12 . The system according to claim 10 , wherein the preprocessor engines are prefetcher engines, and the configurations are different prefetcher engine configurations based on aggressiveness levels to be applied to the prefetcher engines.

13 . The system according to claim 12 , further comprising the prefetcher engines, where in the processor is to:

execute the benchmark applications while changing the aggressiveness levels of the prefetcher engines; and

measure the performance of the device during the execution of the benchmark applications.

14 . The system according to claim 10 , wherein the processor is to:

find a best configuration of the configurations based on the measurements;

find a second-best configuration of the configurations based on the best configuration being found to be the best configuration and based on the measurements; and

find a third-best configuration of the configurations based on the best configuration and the second-best configuration being found to the best configuration and the second-best configuration, respectively, and based on the measurements; and

find an N th best configuration of the configurations based on the best to N−1 best configurations of the configurations being found to be the best to N−1 best configurations, respectively, and based on the measurements.

15 . The system according to claim 14 , wherein the processor is to:

compute first averages of vectors that indicate the performance of the device executing benchmark applications for respective ones of the configurations; and

find a first vector of the vectors having a highest average of the first averages; and

define the best configuration as one of the configurations corresponding to the first vector.

16 . The system according to claim 15 , wherein the processor is to:

(a) select a first candidate vector of the vectors;

(b) perform elementwise maximum operations with the first candidate vector and the first vector yielding a first resultant vector;

(c) compute an average of the first resultant vector;

(d) repeat steps (a) to (c) for other first candidate vectors of the vectors yielding multiple first resultant vectors with multiple corresponding second averages; and

(e) find the second-best configuration corresponding to a second vector of the first candidate vectors yielding a given first resultant vector of the first resultant vectors having a highest average of the second averages.

17 . The system according to claim 16 , wherein the processor is to:

(a) select a second candidate vector of the vectors;

(b) perform elementwise maximum operations with the second candidate vector and the given first resultant vector, or with the first vector and the second vector, yielding a second resultant vector;

(c) compute an average of the second resultant vector;

(d) repeat steps (a) to (c) for other second candidate vectors of the vectors yielding multiple second resultant vectors with multiple corresponding third averages; and

(e) find the third-best configuration corresponding to one of the second candidate vectors yielding one of the second resultant vectors having a highest average of the third averages.

18 . The system according to claim 15 , wherein the processor is to:

(a) select a candidate vector of the vectors;

(b) perform elementwise maximum operations with the candidate vector and at least one vector indicative of the configurations from the best configuration to an N−1 st best configuration, yielding a resultant vector;

(c) compute an average of the resultant vector;

(d) repeat steps (a) to (c) for other candidate vectors of the vectors yielding multiple resultant vectors with multiple corresponding averages; and

(e) find the N th best configuration corresponding to one of the candidate vectors yielding one of the resultant vectors having a highest average of the multiple corresponding averages.

Assignments (1)

Continuity (1)

References Cited (29)