Our benchmark process was aided by a custom listener in the topology which counts the number of tuples entering and leaving the topology and provides a simple static in form of tuples processed per second.
We gathered benchmark data from our large, merged data set:
$ head -50000 taxi_data.csv > benchmark_data.csv
The topologies were started using their respective gradle commands dedicated for benchmarking:
$ gradle bmTopology # benchmark baseline topology
$ gradle bmOptimizedTopology # benchmark optimized topology
After starting the topology we submitted the benchmark data using a submission speed of 1000000:
$ gradle runDataprovider -Pspeed=1000000 -Pdata=./testdata/benchmark_data.csv
Furthermore, in order to isolate the performance from the dashboard we disabled communication between topology and the dashboard for the duration of the benchmark.
We performed the benchmarks on a Intel i7-3517U clocked at 1.90GHz with 4GB RAM.
| Topology | #partitions | parallelism hint | tuples/second |
|---|---|---|---|
| Baseline | 1 | no | 1972.64 |
| Optimized | 1 | no | 3028.46 |
| Optimized | 3 | no | 3553.12 |
| Optimized | 5 | no | 3572.20 |
| Optimized | 3 | 3 | 8542.46 |