Allow benchmark to write json output (#801)

* write json output in benchmark

* fix bugs

* fix

benchmarks/benchmark/tools/profile-generator/container/benchmark_serving.py

@@ -7,6 +7,7 @@
 
 import argparse
 import asyncio
+from datetime import datetime
 import json
 import random
 import time
@@ -266,6 +267,42 @@ async def benchmark(
   await asyncio.gather(*tasks)
 
 
+def save_json_results(args: argparse.Namespace, benchmark_result):
+  # dimensions values are strings
+  dimensions_json = {}
+  # metrics values are numerical
+  metrics_json = {}
+
+  # Setup
+  current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
+  dimensions_json["date"] = current_dt
+  dimensions_json["backend"] = args.backend
+  dimensions_json["model_id"] = args.model
+  dimensions_json["tokenizer_id"] = args.tokenizer
+  if args.additional_metadata_metrics_to_save is not None:
+    dimensions_json = {
+        **dimensions_json,
+        **json.loads(args.additional_metadata_metrics_to_save),
+    }
+  metrics_json["num_prompts"] = args.num_prompts
+
+  # Traffic
+  metrics_json["request_rate"] = args.request_rate
+  metrics_json = {**metrics_json, **benchmark_result}
+
+  final_json = {}
+  final_json["metrics"] = metrics_json
+  final_json["dimensions"] = dimensions_json
+
+  # Save to file
+  base_model_id = args.model.split("/")[-1]
+  file_name = (
+      f"{args.backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"
+  )
+  with open(file_name, "w", encoding="utf-8") as outfile:
+    json.dump(final_json, outfile)
+
+
 def main(args: argparse.Namespace):
   print(args)
   random.seed(args.seed)
@@ -305,24 +342,32 @@ def main(args: argparse.Namespace):
           args.model,
       )
   )
+  benchmark_result = {}
   benchmark_end_time = time.time()
   benchmark_time = benchmark_end_time - benchmark_start_time
   print(f"Total time: {benchmark_time:.2f} s")
   print(f"Requests/min: {60 * args.num_prompts / benchmark_time:.2f}")
+  benchmark_result['benchmark_time'] = benchmark_time
 
   total_output_tokens = np.sum([output_len for _, output_len, _ in
                                 REQUEST_LATENCY])
   output_tokens_per_min = 60 * total_output_tokens / benchmark_time
   print(f"Output_tokens/min: {output_tokens_per_min:.2f}")
+  benchmark_result['total_output_token'] = int(total_output_tokens)
+  benchmark_result['output_tokens_per_min'] = output_tokens_per_min
 
   total_input_tokens = np.sum([prompt_len for prompt_len, _, _ in
                                REQUEST_LATENCY])
   input_tokens_per_min = 60 * total_input_tokens / benchmark_time
   print(f"Input_tokens/min: {input_tokens_per_min:.2f}")
+  benchmark_result['total_input_tokens'] = int(total_input_tokens)
+  benchmark_result['input_tokens_per_min'] = input_tokens_per_min
 
   total_tokens = total_input_tokens + total_output_tokens
   tokens_per_min = 60 * total_tokens / benchmark_time
   print(f"Tokens/min: {tokens_per_min:.2f}")
+  benchmark_result['total_tokens'] = int(total_tokens)
+  benchmark_result['tokens_per_min'] = tokens_per_min
 
   if args.machine_cost:
     print(
@@ -336,6 +381,7 @@ def main(args: argparse.Namespace):
       "Average seconds/request (includes waiting time on server):"
       f" {avg_latency:.2f}"
   )
+  benchmark_result['avg_latency'] = avg_latency
 
   avg_per_token_latency = np.mean([
       latency / (prompt_len + output_len)
@@ -345,6 +391,7 @@ def main(args: argparse.Namespace):
       "Average milliseconds/token (includes waiting time on server):"
       f" {1000 * avg_per_token_latency:.2f}"
   )
+  benchmark_result['avg_per_token_latency'] = avg_per_token_latency
 
   avg_per_output_token_latency = np.mean(
       [latency / output_len for _, output_len, latency in REQUEST_LATENCY]
@@ -353,6 +400,7 @@ def main(args: argparse.Namespace):
       "Average milliseconds/output_token (includes waiting time on server):"
       f" {1000 * avg_per_output_token_latency:.2f}"
   )
+  benchmark_result['avg_per_output_token_latency'] = avg_per_output_token_latency
 
   avg_input_len = np.mean(
       [prompt_len for prompt_len, _, _ in REQUEST_LATENCY]
@@ -361,6 +409,7 @@ def main(args: argparse.Namespace):
       "Average input length:"
       f" {avg_input_len:.2f}"
   )
+  benchmark_result['avg_input_len'] = avg_input_len
 
   avg_output_len = np.mean(
       [output_len for _, output_len, _ in REQUEST_LATENCY]
@@ -369,6 +418,10 @@ def main(args: argparse.Namespace):
       "Average output length:"
       f" {avg_output_len:.2f}"
   )
+  benchmark_result['avg_output_len'] = avg_output_len
+
+  if args.save_json_results:
+    save_json_results(args, benchmark_result)
 
 
 if __name__ == "__main__":
@@ -479,6 +532,18 @@ def main(args: argparse.Namespace):
           " and max_output_length."
       ),
   )
+  parser.add_argument(
+      "--save-json-results",
+      action="store_true",
+      help="Whether to save benchmark results to a json file.",
+  )
+  parser.add_argument(
+      "--additional-metadata-metrics-to-save",
+      type=str,
+      help=(
+          "Additional metadata about the workload. Should be a dictionary in"
+          " the form of a string."
+      ),
+  )
   cmd_args = parser.parse_args()
   main(cmd_args)
-