Performance measurement report for voiceprint-extraction

Hardware

Deployed on kubernetes cluster in Amazon Elastic Kubernetes (EKS)

CPU tests

AWS instance type: c7a.2xlarge
CPU: AMD EPYC 9R14
RAM: 14 GiB

GPU tests

AWS instance type (GPU): g5.2xlarge
CPU: AMD EPYC 7R32
RAM: 14 GiB
GPU: NVIDIA A10G

Software

Image version (CPU tests): 1.7.0
Image version (GPU tests): 1.7.0-gpu
Technology model: xl-5.3.0

Dataset

Name: MO_02
Path: marvin:/_datasets/secure/SIDEvaluationData_01/MO_02/

Test methodology

One instance of the Voiceprint Extraction microservice is deployed on the Kubernetes cluster in EKS.
The client continuously sends randomly selected files to the microservice for 1 hour.
The voiceprint extraction results are processed to extract audio and voice length statistics, then discarded.

⚠️ Warning

Due to a Kubernetes cluster limitation, only 7.7 CPUs are available for the microservice.

CPU

Settings

Test PHX_NUM_THREADS_PER_INSTANCE PHX_NUM_INSTANCES_PER_DEVICE
1 instances, 1 threads 1 1
1 instances, 8 threads 8 1
8 instances, 1 threads 1 8

Results

1 instances, 1 threads 1 instances, 8 threads 8 instances, 1 threads
Processing time [s] 3,602.90 (1h 0m 2.90s) 3,601.22 (1h 0m 1.22s) 3,605.92 (1h 0m 5.92s)
Number of processed files 1,595 5,272 12,682
Total processed audio length [s] 201,832.00 (56h 3m 52.00s) 668,481.00 (185h 41m 21.00s) 1,598,342.00 (443h 59m 2.00s)
Total processed voice length [s] 107,292.95 (29h 48m 12.95s) 352,688.52 (97h 58m 8.52s) 850,913.23 (236h 21m 53.23s)
Faster than realtime (audio) 56.0 185.6 443.3 (55.4 per instance)
Faster than realtime (voiceprint) 29.8 97.9 236.0 (29.5 per instance)
Memory peak [MiB] 290.09 450.98 1424.91

1 instances, 1 threads

1 instances, 8 threads

8 instances, 1 threads

GPU

Settings

Test PHX_NUM_THREADS_PER_INSTANCE PHX_NUM_INSTANCES_PER_DEVICE
1 instances, 1 threads 1 1
8 instances, 1 threads 1 8

Results

1 instances, 1 threads 8 instances, 1 threads
Processing time [s] 3,600.58 (1h 0m 0.58s) 3,601.35 (1h 0m 1.35s)
Number of processed files 11,337 56,520
Total processed audio length [s] 1,427,626.00 (396h 33m 46.00s) 7,107,283.00 (1,974h 14m 43.00s)
Total processed voice length [s] 754,417.51 (209h 33m 37.51s) 3,767,434.03 (1,046h 30m 34.03s)
Faster than realtime (audio) 396.5 1,973.5 (246.7 per instance)
Faster than realtime (voiceprint) 209.5 1,046.1 (130.8 per instance)
Memory peak [MiB] 489.39 699.62
GPU memory peak [MiB] 730 2928

1 instances, 1 threads

8 instances, 1 threads