Automatic and self-optimized determination of execution parameters of a software application on an information processing platform
The invention relates to a method for optimizing the execution parameters of a software application on an information processing platform, consisting in iteratively optimizing said parameters at each execution of said application, in which: for each execution ( 1 ) of said application, performance measurements are made ( 2 ) and stored ( 3 ), in association with the parameters used for said execution; at the start of each execution of said application, values are determined ( 4, 6 ) for a first subset of said parameters by inference ( 6 ) from the stored measurements corresponding to a subset of the executions corresponding to a second subset of said parameters.
1. A method for optimizing parameters for execution of a software application on an information processing platform, the method comprising:
iteratively optimizing said parameters at each execution of said application,
wherein said parameters comprise fixed parameters and free parameters,
wherein for each execution, the method includes, during said execution or once said execution is completed:
making a performance measurement in association with said parameters used for said execution;
creating a stored performance measurement by storing said performance measurement; and
wherein at a start of or right before starting each new execution of said application, determining values for a first subset of said parameters by inference from said stored performance measurement corresponding to a subset of previous executions corresponding to a second subset of said parameters, and
wherein each parameter of the first subset of said parameters is a free parameter and each parameter of the second subset is a fixed parameter, and
wherein said determining comprises inferring optimal values for said first subset of parameters according to a predetermined criterion expressed by said stored performance measurement and from a model describing a relationship between the second and the first subset of parameters, by an optimization method optimizing a cost function, and
wherein said first subset of parameters corresponds to configuration parameters of an input/output accelerator software module for optimizing inputs/outputs of said software application, and
wherein said model is obtained by regression and continuously models a variation in performance of said software application, as a function of the configuration parameters of said input/output accelerator software module.
2. The method according to claim 1 , wherein said second subset is formed by a discriminant parameter such as an application identifier.
3. The method according to claim 1 , wherein said subset of executions is in addition determined by an outlier data filter.
4. The method according to claim 3 , wherein a datum is regarded as an outlier when it concerns a performance measurement value representative of a different behaviour from other values of said performance measurement for said subset of executions.
5. The method according to claim 4 , wherein outlier datum is determined by pairwise comparisons between the data relating to said executions of said subset.
6. A system for optimizing execution parameters of a software application on an information processing platform, comprising:
a processor on an information processing platform;
a non-transitory computer-readable storage medium coupled to the processor, the non-transitory computer-readable storage medium comprising instructions that, when executed, cause the processor to utilize software modules for implementing the method according to claim 1 .
7. A non-transitory computer-readable storage medium comprising instructions for configuring a processor to implement the method according to claim 1 when said instructions are triggered by a data processing platform.