Sky Optimization

Contains documentation about the optimization code used in FogROS2-Sky

module sky_optimization

Functions

terminateClusters(): Terminates all Sky benchmarking instances.

main(): Main Method containing user interface command parser.

OptimizationFunctionType : public Enum

Specifies whether the optimization problem will be solving for time or cost.

Public Static Attributes

TIME = 1: Time Optimization.

COST = 2: Cost Optimization.

OBJECTIVE_COST = 3: Objective Cost Optimization (Motion Planning)

class SkyOptimization

Sky Optimizer responsible for performing the cost and time optimizations based on problem type (Dexnet, VSLAM, Motion Planning) and GPU type (No GPU, T4, etc)

Public Functions

__init__(self, yaml_file, gpu_yaml_file, steps, max_cost, max_time, max_objective_cost, debug, gpu_type, count)

Sky Optimization Constructor.

Parameters:

yaml_file – YAML file specifying Sky benchmarking setup for no GPU
gpu_yaml_file – YAML file specifying Sky benchmarking setup for given GPU type (Tested only on T4)
steps – Number of steps the Sky benchmarking should run
max_cost – User provided cost constraint
max_time – User provided time constraint
max_objective_cost – User provided objective cost constraint
debug – Debug Mode (Lots of print statements)
gpu_type – The GPU type used in the Sky benchmarking (only tested on No GPU, T4)
count – Index in list of SkyOptimization Problems being run (helps coordinate the multithreading)

fullOptimization(self)

Full Sky Optimization Function.

Returns:: List containing Time and Cost Optimizations

getObjectiveCost(self, machine_name)

Parses Sky benchmarking to get objective cost for given machine instance.

Parameters:: machine_name – Name of benchmarking machine instance
Returns:: Objective cost from instance

findInstance(self, optimization_results, optimization_type)

Given results of optimization, this will find the closest feasible hardware setup and its corresponding AWS Instance (Future Work: Build out to other cloud providers)

Parameters:

optimization_results – Scipy Output from optimization
optimization_type – Optimization Type (Cost or Time)

binarySearchRounding(self, optimal_cpu_count)

Rounds CPU output of optimization to nearest feasible CPU amount allowed in cloud instance.

Parameters:: optimal_cpu_count – CPU output from optimization
Returns:: Feasible CPU amount closest to the optimal CPU amount

findGpuValues(self): Parses provided GPU benchmark file to determine possible hardware setups for given GPU.

findCpuValues(self): Parses provided CPU benchmark file to determine edge case hardware setups.

createBenchmarks(self, gpu_type)

Creates Yaml files for each benchmarking instance that will be run.

Parameters:: gpu_type – The GPU type of this optimization instance (only tested on T4 and No GPU)

runSkyBenchmarks(self, count)

Multithreads (Python-style) sky benchmarking instances.

Parameters:: count – Global count relative to other SkyOptimization instances to make sure each Sky benchmarking instance name is unique

skyBenchmark(self, yaml_file, benchmark_name, hardware_count)

Runs sky benchmarking instance from bash script.

Parameters:

yaml_file – Yaml configuration file for benchmarking instance
benchmark_name – Name of Sky benchmarking instance
hardware_count – Tuple containing (CPU, Memory)

regressionSolver(self, data, purpose)

Finds best fitting model given benchmarking data.

Parameters:

data – Benchmarking data for either time, cost, or objective cost
purpose – Time, Cost, or Objective Cost

Returns:

Best fit model and its coefficients

createBounds(self)

Bounds for CPU and Memory for optimization problem.

Returns:: SciPy Bounds object for optimization problem

createConstraints(self, constraint_function_type)

Constraints for Optimization Problem (includes time/cost along with memory/CPU ratio)

Parameters:: constraint_function_type – Constraint function type, which is opposite of the optimization problem (Time Optimization -> Cost Constraint and Cost Optimization -> Time Constraint)
Returns:: SciPy Constraints object for optimization problem

getTimeOptimization(self)

Gets Time Optimization Results.

Returns:: Time Optimization Results

getCostOptimization(self)

Gets Cost Optimization Results.

Returns:: Cost Optimization Results

createObjectiveFunction(self, objective_function_type)

Optimization Loss Function.

Parameters:: objective_function_type – Time or Cost
Returns:: Optimization Loss Function

memoryFunction(self)

Function for memory/CPU ratio constraint.

Returns:: Memory:CPU Constraint Function

costFunction(self)

Cost Function = $/second * seconds/step * # steps.

Returns:: Function object for the Cost Function

timeFunction(self)

Time Function = seconds/step * # steps.

Returns:: Function object for the Time Function

objectiveCostFunction(self)

Objective Cost Function.

Returns:: Function object for the Objective Cost Function

solveOptimization(self, objective_function_type, constraint_function_type)

Optimization Problem Setup and Solver.

Parameters:

objective_function_type – Time or Cost
constraint_function_type – Opposite of objective_function_type

Returns:

Optimization Result, Time Model Result, Cost Model Result

Public Members

debug_: Debug Mode (Lots of print statements)

steps_: Number of steps the Sky benchmarking should run.

max_cost_: User provided cost constraint.

max_time_: User provided time constraint.

max_objective_cost_: User provided objective cost constraint.

gpu_type_: The GPU type used in the Sky benchmarking (only tested on No GPU, T4)

count_: Index in list of SkyOptimization Problems being run (helps coordinate the multithreading)

SECONDS_PER_MINUTE: Number of seconds in a minute.

MINUTES_PER_HOUR: Number of minutes in an hour.

SECONDS_PER_HOUR: Number of seconds in an hour.

ZERO_EPSILON: Positive zero approximation.

yaml_file_: YAML file specifying Sky benchmarking setup for no GPU.

gpu_yaml_file_: YAML file specifying Sky benchmarking setup for given GPU type (Tested only on T4)

is_slam_: Boolean indicating if the task is V-SLAM.

is_motion_planning_: Boolean indicating if the task is motion planning.

is_dexnet_: Boolean indicating if the task is Dexnet.

optimization_function_type_length_: Index in solution string specifying optimization function type.

max_cpu_: Initialzation of maximum CPU constraint.

max_memory_: Initialization of maximum memory constraint.

relevant_cpu_list_: List of realistic CPU setups for the cloud (similar to relevant hardware list without memory, rememant of prototype version that just looked for optimal CPU)

relevant_hardware_list_: List of realistic CPU and memory setups for the cloud.

benchmark_time_results_: List storing results from time benchmarking.

benchmark_yaml_list_: List of YAML files generated that will be used for Sky benchmarking.

benchmark_range_: Number of benchmarks performed per feature (i.e we only do benchmarking for minimum CPU amount, maximum CPU amount, and middle CPU amount)

benchmark_threads_: List of different threads running Sky benchmarking.

benchmark_cost_substring_1_: First identifier substring for cost benchmarking.

benchmark_cost_substring_2_: Second identifier substring for cost benchmarking.

benchmark_time_substring_1_: First identifier substring for time benchmarking.

benchmark_time_substring_2_: Second identifier substring for time benchmarking.

min_gpu_substring_1_: First identifier substring for GPU identification.

min_gpu_substring_2_: Second identifier substring for GPU identification.

mem_to_cpu_ratio_for_gpu_: Initialization for memory:CPU ratio (used only when GPU type isn’t None)

num_steps_and_duration_: List containing step and startup time results for benchmarking (remmanat of prototype version where we tried to account for startup time.

average_setup_time_: Average setup time for cloud instance (remmenant of prototype version: see self.num_steps_and_duration_ for explanation)

got_cost_info_: Boolean indicating if we have receieved cost information from benchmarking.

benchmark_cost_substring_cpu_index_: Indiex where CPU information is located in benchmark cost string.

benchmark_cost_substring_cost_index_: Index where Cost information is located in benchmark cost string.

benchmark_cost_results_: List containng cost results from Sky benchmarking.

benchmark_objective_cost_results_: List containing objective cost results for Sky benchmarking (only relevant for Motion Planning)

yaml_file_contents_: String containing YAML file contents for Sky benchmark YAML.

cpu_start_index_: Index containing information about where CPU information starts in YAML file.

cpu_candiate_index_: Index containing information about where CPU candidate information starts in YAML file.

cpu_str_offset_: Offset for index where we insert CPU information into the YAML file.

seconds_per_step_index_: Index containing seconds per step information during Sky benchmarking.

num_steps_index_: Index containing number of steps information during Sky benchmarking.

duration_min_index_: Index containing minute duration during Sky benchmarking.

duration_sec_index_: Index containing second duration during Sky benchmarking.

solutions_in_constraints_: List of solutions that are within constraints.

solutions_outside_constraints_: List of solutions that are outside constraints.

data_file_: String before relevant benchmarking data (Let’s the program know which line to collect benchmarking data from)

data_file_name_: Data file we will save benchmark data results.

result_queue_: List containing info that will be outputted to data file.

possible_cpus_: List containing possible CPUs (current cloud setup could not do any other CPU amounts)

optimal_cost_instance_: String stating the optimal cloud instance for cost optimization.

optimal_time_instance_: String stating the optimal cloud instance for time optimization.

optimal_objective_cost_instance_: String stating the optimal cloud instance for objective cost optimization.

t4_map_: Map between hardware tuples (CPU, Memory) and T4 GPU cloud instances for AWS (Future Work: Will generate these maps online by parsing SkyPilot database CSV file)

no_gpu_map_: Map between hardware tuples (CPU, Memory) and no GPU cloud instance (Future Work: Will generate these maps online by parsing SkyPilot database CSV file)

time_model_info_: Time Regression Output containing model info and constraints.

cost_model_info_: Cost Regression Output containing model info and constraints.

objective_cost_model_info_: Objective Cost Regression Output containing model info and constraints (only motion planning)

timeModel_: Time Regression Model.

time_model_coefficients_: Coefficients for Time Regression Model.

costModel_: Cost Regression Model.

cost_model_coefficients_: Coefficients for Cost Regression Model.

objectiveCostModel_: Objective Cost Regression Model.

objective_cost_model_coefficients_: Coefficients for Objective Cost Regression Model.

cost_optimization_results_: Scipy Output from Cost Optimization.

time_optimization_results_: Scipy Output from Time Optimization.

objective_cost_optimization_results_: Scipy Output from Objective Cost Optimization.