Machine learning based contention delay prediction in multicore architectures
A method of generating training data for training a Machine Learning based Task Contention Model, ML based TCM, to predict time delays resulting from contention between tasks running in parallel on a multi-processor system is provided herein. The method includes: executing a plurality of microbenchmarks, μBenchmarks B j , on the multi-processor system in isolation and measuring a number of resultant Performance Monitoring Counters, PMCs, over time to extract ideal characteristic footprints of each μBenchmark when operating in isolation; performing a feature correlation analysis on the PMCs resulting from the plurality of μBenchmarks to determine the degree of correlation between each resultant PMCs and the executed plurality of μBenchmarks; selecting a number of PMCs based upon their degree of correlation between the plurality of μBenchmarks to form a reduced PMC array.
1 . A method of generating training data for training a Machine Learning based Task Contention Model (ML based TCM), to predict time delays resulting from contention between tasks running in parallel on a multi-processor system, the method comprising:
executing a plurality of microbenchmarks, microbenchmarks B j , on the multi-processor system in isolation and measuring a number of resultant Performance Monitoring Counters, PMCs, over time to extract ideal characteristic footprints of each microbenchmark when operating in isolation;
performing a feature correlation analysis on the PMC metrics resulting from the plurality of microbenchmarks to determine the degree of correlation between each resultant PMCs and the executed plurality of microbenchmarks; and
selecting a subset of PMCs based upon their degree of correlation between the plurality of microbenchmarks to form a reduced PMC array,
wherein the multi-processor system is a multi-core processor of an avionics system, and wherein the multi-processor system is a homogeneous platform, a heterogeneous platform or asymmetric.
2 . The method of claim 1 , wherein the plurality of microbenchmarks are selected based on the Arithmetic Intensity, AI, of each microbenchmark to stress certain interference channels of the multi-processor system in an isolated way.
3 . The method of claim 1 , wherein the plurality of microbenchmarks are selected from a pre-populated code block repository, and are selected so as to generate the desired interference and contention scenarios for training data that may be used in training an accurate ML-based TCM.
4 . The method of claim 1 , wherein the plurality of microbenchmarks are selected such that they each individually have a shorter execution time individually than a maximum makespan for a given task to be scheduled on the multi-processor system.
5 . A computer system for producing training data for training a Machine Learning based Task Contention Model (ML based TCM) to predict time delays resulting from contention between tasks running in parallel on a multi-processor system, wherein the computer system is configured to perform the method claim 1 .
6 . A non-transitory computer-readable medium storing instructions which, when executed on a computer system, cause the computer system to produce training data for training a Machine Learning based Task Contention Model, ML based TCM, to predict time delays resulting from contention between tasks running in parallel on a multi-processor system, by performing the method of claim 1 .
7 . A computer-implemented method of producing a trained Machine Learning based Task Contention Model (ML based TCM), to predict time delays resulting from contention between tasks running in parallel on a multi-processor system using training data generated by the method of any preceding claim , the method comprising:
executing a predetermined set of possible pairing scenarios of the plurality of microbenchmarks in parallel on the multi-processor system and measuring the effect on the execution time of each microbenchmark, ΔT B j , resulting from contention over interference channels within the multi-processor system; and
training a machine learning model using, as an input, a reduced PMC array for each microbenchmark in isolation and, at the output, the corresponding ΔT B j during the parallel execution of each pairing scenario as training inputs,
wherein the multi-processor system is a multi-core processor of an avionics system, and wherein the multi-processor system is a homogeneous platform, a heterogeneous platform or asymmetric.
8 . The computer-implemented method of claim 7 , wherein the machine learning model is a decision tree-based predictor, and wherein the machine learning model is an XGBoost model.
9 . The computer-implemented method of claim 7 , comprising validating the training error of the ML based TCM by:
executing a plurality of actual execution tasks on the multi-processor system in isolation and measuring the resultant PMCs corresponding to the reduced PMC array over time for each actual execution task;
inferring, by the ML based TCM, the predicted effect on the execution time of each actual execution task at least the resultant PMCs of each task as input;
executing the actual execution tasks in parallel and measuring the actual execution time; and
comparing the predicted effect on the execution time with the actual execution time, thereby calculating an error between the predicted and actual execution time.
10 . The computer-implemented method of claim 9 , wherein the training of the machine learning model comprises:
a first iterative loop over all different pairing scenarios; and
a second iterative loop over each of the selected PMC measures of each microbenchmark in isolation, as well as the corresponding ΔT B j during the parallel execution of each pairing scenario.
11 . The computer implemented method of claim 7 , wherein the measuring the selected PMCs comprises measuring the selected PMCs at a variable monitoring frequency.
12 . A computer system comprising:
at least one processor; and
a memory storing a Machine Learning based Task Contention Model configured to predict time delays resulting from contention between tasks running in parallel on a multi-processor system,
wherein the Machine Learning based Task Contention Model is produced by the method of claim 7 .
13 . A computer system for producing a trained Machine Learning based Task Contention Model (ML based TCM), to predict time delays resulting from contention between tasks running in parallel on a multi-processor system, wherein the computer system is configured to perform the method of claim 7 .
14 . A non-transitory computer-readable medium storing instructions which, when executed on a computer system, cause the computer system to produce a trained Machine Learning based Task Contention Model (ML based TCM) to predict time delays resulting from contention between tasks running in parallel on a multi-processor system, by performing the method of claim 7 .