Optimize step

An optimize step generates optimized configurations according to the defined optimization strategy. During this step, Akamas AI is used to generate such optimized configurations.

The optimize step has the following structure:

Optimizer

The optimizer field allows selecting the desired optimizer:

• AKAMAS identifies the standard AI optimizer used by Akamas

• SOBOL identifies an optimizer that generates configurations using Sobol sequences

• RANDOM identifies an optimization that generates configurations using random numbers

Notice that SOBOL and RANDOM optimizers do not perform initialization experiments, hence the field numberOfInitExperiments is ignored.

Optimizer options for offline studies

For offline optimization studies, the optimizerOptions field can be used to specify whether beta-warping optimization (a more sophisticated optimization that requires longer time) should be used for how many experiments (as a percentage):

# half the experiments
# should be done
# with beta warping
experimentsWithBeta: "50%"

where experimentsWithBeta can be:

• A percentage between 0 and 100%

• A number less than or equal to numberOfExperiments

Please notice that also the safetyFactor option discussed here below in the context of live optimization studies can be applied to offline optimization studies.

Optimizer options for live studies

For live optimization studies, the optimizerOptions field can be used to specify several important parameters governing the live optimization, which can be defined at the study level and also overridden at the step level (only for steps of type optimize):

optimizerOptions:
  onlineMode: RECOMMEND                    # [RECOMMEND|FULLY_AUTONOMOUS]
  safetyMode: GLOBAL                       # [GLOBAL|LOCAL]
  workloadOptimizedForStrategy: MAXIMIN    # [MAXIMIN|MEDIAN|LAST|MOST_VIOLATED]
  safetyFactor: 0.55                       # 0 <= safetyFactor <= 1
  explorationFactor: 0.05                  # 0 <= explorationFactor <= 1 or FULL_EXPLORATION

Notice that while available as independent options, the optimizer options onlineMode (described here below), workloadOptimizedForStrategy (here below) and the safetyFactor (here below) works in conjunction according to the following schema:

All these optimizer options can be changed at any time, that is while the optimization study is running, to become immediately effective. The Optimizer options commands page in the reference guide provides these specific update commands.

Online Mode

The onlineMode field specifies how the Akamas optimizer should operate:

• RECOMMEND: configurations are recommended to the user by Akamas and are only applied after having been approved (and possibly modified) by the user;

• FULLY AUTONOMOUS MODE: configurations are immediately applied by Akamas.

Safety Mode

The safetyMode field describes how the Akamas optimizer should evaluate the goal constraints on a candidate configuration for that configuration to be considered valid:

• GLOBAL: the constraints must be satisfied by the configuration under all observed workloads in the configuration history - this is the value taken in case onlineMode is set to RECOMMEND;

• LOCAL: the constraints are evaluated only under the workload selected according to the workload strategy - this should be used with onlineMode set to FULLY_AUTONOMOUS.

Notice that when setting the safetyMode to LOCAL, the recommended configuration is only expected to be good for the specific workload selected under the defined workload strategy, but it might violate constraints under another workload.

Workload Strategy

The workloadOptimizedForStrategy field specifies the workload strategy that drives how Akamas leverages the workload information when looking for the next configuration:

• MAXIMIN: the optimizer looks for a configuration that maximizes the minimum improvements for all the already observed workloads;

• MEDIAN: for each workload, a median of all its values is considered - this works well to find a configuration that is good for the median of all the workloads;

• LAST: for each workload, the last observed workload is considered - this works well to find a configuration that is good for the last workloads - this is often used in conjunction with a LOCAL safety mode (see here above);

• MOST_VIOLATED: for each workload, the workload of the configuration which results in most violations is considered.

Safety Factor

The safetyFactor field specifies how much the optimizer should stay on the safe side in evaluating a candidate configuration with respect to the goal constraints. A higher safety factor corresponds to a safer configuration, that is a configuration that is less likely to violate goal constraints.

Acceptable values are all the real values ranging between 0 and 1, with (safetyFactor - 0.5) representing the allowed margin for staying within the defined constraint:

• 0 means "no safety", as with this value the optimizer totally ignores goal constraints violations;

• 0.5 means "safe, but no margin", as with this value the optimizer only tries configurations that do not violate the goal constraints, by remaining as close as possible to them;

• 1 means "super safe", as with this value the optimize only tries configurations that are very far from goal constraints.

For live optimization studies, 0.6 is the default value, while for offline optimization studies, the default value is 0.5.

Exploration Factor

The explorationFactor field specifies how much the optimizer explores the (unknown) optimization space when looking for new configurations. For any parameter, this factor measures the delta between already tried values and the value of a new possible configuration. A higher exploration factor corresponds to a broader exploration of never tried before parameter values.

Acceptable values are all the real values ranging between 0 and 1, plus the special string FULL_EXPLORATION:

• 0 means "no exploration", as with this value the optimizer chooses a value among the previously seen values for each parameter;

• 1 means "full exploration, except for categories", as with this value the optimizer for a non-categorical parameter any value among all its domain values can be chosen, while only values (categories) that have already been seen in previous configurations are chosen for a categorical parameter;

• FULL_EXPLORATION means "full exploration, including categories" as with this value the optimizer chooses any value among all its domain values, including categories, even if not already seen in previous configurations.

In case the desired explorationFactoris 1 but there are some specific parameters that also need to be explored with respect to all its categories, then PRESET steps (refer to the Preset step page) can be used to run an optimization study with these values. For an example of a live optimization study where this approach is adopted see Optimizing a live full-stack deployment (K8s + JVM).

Failures

An optimize step is fault-tolerant and tries to relaunch experiments when they fail. Nevertheless, the step imposes a limit on the number of failed experiments: if too many experiments fail, then the entire step fails too. By default, at most 30 experiments can fail while Akamas is optimizing systems. An experiment is considered failed when it either failed to run (i.e., there is an error in the workflow) or violated some constraint.

Inizialitations

An optimize step launches some initialization experiments (by default 10) that do not apply the AI optimizer and are used to find good configurations. By default, the step performs 10 initialization experiments.

Initialization experiments take into account bootstrapped experiments, experiments executed in preset steps, and baseline experiments.

Examples

The following fragment refers to an optimization study that runs 50 experiments using the SOBOL optimizer:

name: "my_optimize" # name of the step
type: "optimize" # type of the step (optimize)
optimizer: "SOBOL"
numberOfExperiments: 50 # amount of experiments to execute
numberOfTrials: 2 # amount of trials for each experiment

The following fragment refers to an optimization where 50% of the experiments need to use the beta-warping option, which enables a more sophisticated but longer optimization:

# half the experiments
# should be done
# with beta warping
experimentsWithBeta: "50%"