[Question] About perf_analyzer request & execution count

Hello guys. I have a question about perf_analyzer request & execution count.

root@ai-gpu-2:/workspace# perf_analyzer -m ensemble_models -b 32 --shape INPUT:640,640,3 
*** Measurement Settings ***
  Batch size: 32
  Using "time_windows" mode for stabilization
  Measurement window: 5000 msec
  Using synchronous calls for inference
  Stabilizing using average latency

Request concurrency: 1
  Client: 
    Request count: 17
    Throughput: 108.8 infer/sec
 ...
  Server: 
    Inference count: 672
    Execution count: 21
    Successful request count: 21

Why is execution count > request count in the example below? What exactly does execution count mean? Thanks in advance.

[GuanLuo]

Execution count is the number of the model run, where request count is the number of inference request sent to the model. When using dynamic batching, Triton may batch multiple requests for one model run and in such case you can have smaller execution count than request count.

[GuanLuo]

The client stats and server stats are being collected in different location where client stats is collected inside perf analyzer while server stats is calculated based on the delta reported from server stats API, so the variance will be observed if you compare client stats with server stats.