• Online Graph summarisation experiment: https://github.com/AU-DIS/A-GLIMPSE-Reloaded
• Index tuning experiment: https://github.com/AU-DIS/DBABandits-modded and https://github.com/Koenkalle/HMAB-modded/tree/main

A modified version of https://github.com/malingaperera/DBABandits with the addition of QBL

The setup requires a connected mssql TPC-H server to run. We provide an overview of a setup using microsofts provided docker image.

Pull mssql container with
sudo docker pull mopt/mssql-tools/bin/scr.microsoft.com/mssql/server:2022-latest

Run it with

sudo docker run -e "ACCEPT_EULA=Y" -e "MSSQL_SA_PASSWORD=L0ngpassword!" \
-it --mount src="~/path/to/schema",target=/mnt,type=bind \
   -p 1433:1433 --name sql1 --hostname sql1 \
   -d \
   mcr.microsoft.com/mssql/server:2022-latest

Which mounts a path that can be used to save table files into. Remember to set a suitable password.

Create schema files with TPC-H generator. One way is outlined here: https://github.com/nghiahhnguyen/TPC-H-Dataset-Generator-MS-SQL-Server It may require minor modification of the bulkloading and .ddl and .ri files depending on setup.

An easy way to access the is with mssql-cli

You can then connect with mssql-cli -S localhost -U SA -P L0ngpassword! and run sql commands e.g. DROP DATABASE tpc and CREATE DATABASE tpc

With a database created you can run files such as mssql-cli -S localhost -U SA -P L0ngpassword! -d tpc -i dss.ddl & mssql-cli -S localhost -U SA -P L0ngpassword! -d tpc -i dss.ri

Which can reference schema table files from the mounted /mnt folder

This repository contains experimental setups to compare bandit approaches against Glimpse for online knowledgegraph summarisation.

1. Install Python packages from requiriements.txt

2. Download dbpedia39

   2.1 run downLoadDBPedia39() from util.py to download the requiried rdf files into folder dbpedia/rdf_gz/

   2.3 outcomment downLoadDBPedia39() and uncomment the graph load and save instead. Run it.

   2.2 run subgraphs.py to create a simple smaller test graph for quick experimentation. Its generation can freely be modified.

   2.3 if any step fail. You may have to update the folder path to dbpedia/rdf_gz in both glimpse/src/experiment_base.py and glimpse/src/base.py for class DBPedia(KnowledgeGraph):.

3. run main.py and check if it works!

4. Plots are saved in replacements_results/. Beware that a lot of naming is legacy stats that may or may not mean anything.

There is 4 things of interest to modify.

• The methods
• The parameters
• The queries
• The graphs

The Methods

Bandits are saved ins bandits/effecients_bandits/. The Exp3 variants are older implementations that I may update. Changing bandits is currently not that flexible and has to be done manually in run_compare_experiment() line ~60 by changing between bandit="exp3" and bandit="qbl". To make a new bandit simply create a new file and implement the same methods. Then add it in glimpseonline.py like the other bandits.

Note All of this method swapping can be automated.

The Parameters

Bandit parameters are currently set individually and very unorganised.

Most Experiment parameters can be found in main.py. You can execute multiple experiments at once multi threaded (See below). For now use only run_compares() as other methods contain legacy.

graph: name of graph to load.
graph_size: custom defined size of the graph (LEGACY when using run_compares())
n: number of timesteps.
ks: target size of summary graph proportional to size of the full graph.
batch_size: number of queries provided in each timestep.
query_generator: name of query generator to use.

ks, batch_size, query_Generator is wrapped in a list, allowing to execute experiments with multiple settings at once.

In experiments/bandit_versus_glimpse.py the variable number_of_aggregations can be set to make an average result of each experiment. number_of_aggregations=1 executes each experiment once. number_of_aggregations=10 executes each experiment 10 times and report the average.

The graph

All experiments are based on dbpedia or subgraphs thereof. subgraphs can be generated. See examples of how in subgraphs.py. Things to consider: to generate some more meaningfull graphs, it would be smart to utilize graph traversal, otherwise they become very disconected. Though that is fine in some cases e.g. special query loads or just for simple code experimentation.

To run multiple experiments in parallelle, simply uncomment

                    #p = Process(target=run_compare_experiment,
                    #            args=(graph, n, k, bs, rf, query_generator))
                    #processes.append(p)

    #for p in processes:
        #Too much parralell if splitting this
        #p.start()
        #p.join()

in run_compares() and remove the old run_compare_experiment() line above it.

• https://github.com/vincentlaucsb/csv-parser?ref=bestofcpp.com#integration
• Google benchmark
• boost