@@ -266,7 +266,7 @@ For more information see [Collecting Performance Numbers](#collecting-performanc
## Collecting Performance Numbers
Currently, there is no consistent way of obtaining the configuration of deployment service. Hence, we need to provide this information to the script in form of command line arguments. The benchmarking script `/workspace/examples/llm/benchmarks/perf.sh` uses GenAI-Perf tool to collect the performance numbers at various different request concurrencies. The perf.sh script can be run multiple times to collect numbers for various different deployments. Each script execution will create a new artifacts directory in `artifacts_root` and dump these numbers in it. See [Plotting Pareto Graphs](#plotting-pareto-graphs) to learn how to convert the data from this `artifacts_root` to generate pareto graphs for the performance.
Currently, there is no consistent way of obtaining the configuration of deployment service. Hence, we need to provide this information to the script in form of command line arguments. The benchmarking script `/workspace/benchmarks/llm/perf.sh` uses GenAI-Perf tool to collect the performance numbers at various different request concurrencies. The perf.sh script can be run multiple times to collect numbers for various different deployments. Each script execution will create a new artifacts directory in `artifacts_root` and dump these numbers in it. See [Plotting Pareto Graphs](#plotting-pareto-graphs) to learn how to convert the data from this `artifacts_root` to generate pareto graphs for the performance.
Note: As each `perf.sh` adds a new artifacts directory in the `artifacts_root` always, proper care should be taken that we are starting experiment with clean `artifacts_root` so we include only results from runs that we want to compare.
