TPGenerator: Lauca

• The code in TPGenerator provides the component of Transaction Logic Analyzer, Data Access Distribution Analyzer, Performance Monitor, Data Characteristics Extractor, Workload Generator and Database Generator in code.
• Technique Report can be found in technical report.

Quick Start for TiDB with PingCAP

• We can simply use the following commands to simulate data and workloads. The TiDB example in LaucaExample. Here, we provides a configuration file about TiDB in both LaucaProduction and LaucaTesting，and according jar files in LaucaProduction.jar and LaucaTesting.jar.

cd LaucaExample
cd Production
# this step will automatically catpured the data characteristics and then create dataCharacteristicSaveFile.obj in testdata file
java -jar LaucaProduction.jar ./lauca-tidb.conf --getDataCharacteristics
# this step will automatically catpured the workload characteristics(transaction logic and data access distribution) and then create txLogicSaveFile.obj  and distributionSaveFile.obj in testdata file
java -jar LaucaProduction.jar ./lauca-tidb.conf --getWorkloadCharacteristics

#open anothor terminal
cd LaucaExample
cd Testing
# this step will automatically generate data in testdatab/data according to the dataCharacteristicSaveFile.obj
java -jar LaucaTesting.jar ./lauca-tidb.conf --geneSyntheticDatabase
# this step will automatically load data from testdatab/data into TiDB database which is configurated in lauca-tidb.conf.
java -jar LaucaTesting.jar ./lauca-tidb.conf --loadSyntheticDatabase
# this step will automatically generate workload into the TiDB database according dataCharacteristicSaveFile.obj, txLogicSaveFile.obj  and distributionSaveFile.obj.
java -jar LaucaTesting.jar ./lauca-tidb.conf --geneSyntheticWorkload

Example to get workload trace

• We can get the workload trace via skywalking, the link is https://github.com/apache/skywalking
• The application can print workload trace through java agent. Here we use oltp-bench which generates TPC-C as our application side. User can use the following step to load database and print workload trace.

cd oltpbench
# configurate the database info
vim config/tpcc_config_tidb.xml
# load the database
./oltpbenchmark -b tpcc -c config/tpcc_config_tidb.xml --create=true --load=true
# generate the workload into databases and print the workload trace
./oltpbenchmarkSkywalking -b tpcc -c config/tpcc_config_tidb.xml --execute=true

• But we also provide a workload trace about TPC-C with 20 warehouses, 20 connections and 100 second running time available in workload trace

Overview of Program Files

• In the LaucaExample folder, there are three sub folders, namely, the production environment, the testing environment and testdata. In the production environment, the basic data feature dataCharacteristicSaveFile.obj, the transaction logic feature txLogicSaveFile.obj and the workload access distribution feature distributionSaveFile.obj are generated in testdata. The testing environment generates a synthetic database using the basic data features, which create tables file in testdata and then used to generate database through its second step. In addition, the testing environment generates simulated workloads using transaction logic and workload access distribution. LaucaProduction.jar is in the Production folder and LaucaTesting.jar is in the Testing folder. The former one is responsible for extracting the data features and workloads features, while the latter oen is responsible for generating the simulated database and simulated workloads.

LaucaExample/
├── Production
│   ├── lauca-tidb.conf
│   ├── LaucaProduction.jar
│   └── log4j.properties
├── testdata
│   ├── dataCharacteristicSaveFile.obj
│   ├── distributionSaveFile.obj
│   ├── tables
│   └── txLogicSaveFile.obj
└── Testing
    ├── lauca-tidb.conf
    ├── LaucaTesting.jar
    └── log4j.properties

Supplementary Experiments

Experiment Setting

• We conduct our experiments on a sever equipped with 48 Intel(R) Xeon(R) Gold 6126 @ 2.60Ghz CPUs, 250GB memory, 8TB disk.
• We deploy both client and database service on the same machine.
• We use MySQL v5.7.27 and TiDB v5.0.0.

Experiment for Review1-O2

• In order to illustrate that our simulation is also good if the production environment and the evaluation environment use different databases, we run a new set of experiment.
• Supposing we load TPC-C workload into MySQL, Lauca extracts the characteristcis of both data and workload to generate a simulated scenario (data and workload). The simulated scenario and the real scenario (TPC-C workload from OLTP-Bench) are loaded into TiDB, respectively. We then compare their performance on TiDB.
• Figure 1 shows performance deviations between simulated workloads generated by Lauca and real workloads generated by OLTP-Bench, under 10 warehouses and 10 threads. We can see that these two workloads are very similar in throughput and average latency, where the deviations are 2.94% and 6.65% respectively.

Figure 1

### Experiment for Review2-O6

• We deploy the client and server on the same machine through the isolation of NUMA node.
• We do experiments respectively through local loopback and unix domain socket to avoid the slow network.
• In this experiment, we run TPC-C workload on MySQL (v5.7.27), using a database of 20 warehouses.
• Figure 2(a)-(c) shows performance deviations between simulated workloads generated by Lauca and real workloads generated by OLTP-Bench.

​ Figure 2(a) Throughput Figure 2(b) Avg Latency Figure 2(c) 95% Latency

Example for calculate probability of a dependency relationship (Review2-O1(3))

• Here we take an equal relationship (ER) calculation as an example.
• Suppose TXN₁ has two statements:

  update S set s₃ = s₃ + p_1,1 where s₁ = p_1,2; update T set t₃ = t₃ + p_2,1 where t₁ = p_2,2 and t₂ = p_2,3;

• If the total number of transaction instances for TXN₁ is T. When calculating ER probability for p_2,3 , we compare it with the values from each previous statement, i.e., p_1,1, p_1,2, p_2,1 and p_2,2. If p_1,2 = p_2,3 in t instances, we have ER for p_2,3, i..e, [p_1,2, p_2,3, ER, t/T].