SketchGuide

This is github repository for SketchGuide.

Quick Start

You can first modify the path and reconfigure a CM-sketch optimized by SketchLib O6 using SketchGuide.

Step 1

Install bayesian-optimization

$ pip install bayesian-optimization

Step 2

You should modify these paths in the following files in your environment.

• ./cmO6/p4src/compile_cm_O6.sh
  • dir = YOUR_DIR/open_src/
  • script_home = SDE_DIR
• ./cmO6/perform_mapper/resource_mapper.py
  • dir = "YOUR_DIR/open_src/cmO6"
  • resource_log_file = "/SDE_DIR/build/p4-build/p416_countmin_O6/tofino/p416_countmin_O6/pipe/logs/mau.resources.log"
  • compile_sh = "YOUR_DIR/open_src/cmO6/P4src/compile_cm_O6.sh"

Step 3

Then, you can run the supervisor file to automatically generate sketch skyline candidates

$ python3 supervisor.py

Step 4

Last, you can solve the final sketch skyline results

$ python3 skylines.py

You can see file "skyline_results.log" is generated.

Tutorial

• You can leverage SketchGuide to reconfigure your sketches.

Step 1

• Write a P4 program that can be successfully compiled. e.g., ./p4src/p416_countmin_O6.p4.
• Write a sketch program that can output the accuracy results, e.g., ./perform_mapper/accuracy_mapper.py.

Step 2

• Identify the resource parameters (referring to the primitives in our paper).
• Replace the parameters that you want to configure as a stub, and generate a P4 template file, e.g., ./p4src/template.p4
• Modify the ./p4src/stub.py, which can identify the stubs and replace the stubs as new parameters so that a completed P4 program can be generated.

Step 3

• Set the parameter range of sketches in the ./config/config.json, the parameter range can be relatively large because the Bayesian Optimizer will solve and narrow down the range later.
• Set the weight of each metric that your want to optimize. The Bayesian Optimizer will combine these metrics as a linear combination and maximize this combination.
• Set the bound of results. You can budget the resources and filter the unsatisfactory results here.

Step 4

• Modify the input parameter of black_box_function in ./supervisor.py, e.g., def black_box_function(reg_size, hash_cnt, hash_tbl_size, exact_tbl_size)
• Modify the accuracy mapper for your parameters, e.g., recall, precision = am.accuracy_mapper(params['reg_size'], params['hash_cnt'], params['hash_tbl_size'], params['exact_tbl_size'])
• Keep the same parameter order as the one in ./config/config.json
• Adjust the maximal metric value in fuction get_max_metric_val (supervisor.py) to normalized the objective function.
• Choose the rounds of bayesian optimization in function optimized_params = solve_candidate(init_points = 10, n_iter = 40), 10 randomly search.

Step 5

• Run supervisor.py to generate SS candidates.
• Run skylines.py to solve the final SS results.