Edge workload traces from Aeneas, Julius, and MR-Leo

Overview

The emerging field of edge computing is suffering from a lack of representative data to evaluate rapidly emerging new algorithms or techniques. It is a critical issue as this complex paradigm has numerous different use cases which translate to a highly diverse set of workload types.

In this dataset, we collect traces for three available edge applications: Aeneas [1], Julius [2], and MR-Leo [3]. They are provided open-source.  Read our article [4] for explanations about the gathering of these traces and an analysis of them.

[1] https://www.readbeyond.it/aeneas/
[2] https://github.com/julius-speech/julius
[3] https://gitlab.liu.se/ida-rtslab/public-code/2019_mrleo_server
[4] K. Toczé, N. Schmitt, U. Kargén, A. Aral, and I. Brandic, Edge Workload Trace Gathering and Analysis for Benchmarking, in 6th IEEE International Conference on Fog and Edge Computing 2022 (ICFEC 2022), IEEE, 2022. DOI: 10.1109/ICFEC54809.2022.00012

Details about the hardware used for the trace collection

The Aeneas trace was gathered using a an HP Elitebook 840 G5 running Ubuntu 18.10. It has 16 GB RAM and an Intel Core i7-8550U CPU (1.8 GHz, 4 cores, 8 threads). Aeneas version 1.7.3 was used.

We collect traces for Julius on four Off-The-Shelve servers with varying performance as depicted in the following table. Because Julius is a single-core application, performance is mostly dictated by the clock speed of the CPU. Each server is running Ubuntu 18.04.4 LTS with kernel version 4.15.0-108-generic.
As some audio files caused Julius to crash, all files are converted from the original Ogg Vorbis format to WAV by applying \texttt{ffmpeg -i audio.ogg -acodec pcm\_s16le -ac 1 -ar 16000}, resulting in a total of 5724 audio files.

System under test servers for the Julius speech recognition application. (All CPUs are Intel Xeon)

Server
CPU @ Clock (Cores / Threads)
Memory @ Clock

A
E3-1230 v5 @ 3.4GHz (4/8)
1x16GB @ 2133MHz

B
E5-2640 v3 @ 2.6GHz (8/16)
2x16GB @ 2133MHz

C
E5-2650 v3 @ 2.3GHz (10/20)
2x16GB @ 2133MHz

D
E5-2650 v4 @ 2.2GHz (12/24)
2x16GB @ 2133MHz

The MR-Leo trace is gathered using a an HP Elitebook 840 G5 running Ubuntu 18.10. It has 16 GB RAM and an Intel Core i7-8550U CPU (1.8 GHz, 4 cores, 8 threads). The MR-Leo implementation using ORB-SLAM2 is used, in the replay mode on the edge server, meaning that the input video was directly streamed on the edge server, and not from an end device.

Details about the task id field

The task id for the Aeneas trace looks as follows: A_1_3 where the A stands for Aeneas, the first number corresponds to the example used for this task (see Figure 3 in parenthesis) and the second number identify the run number in which the data is collected.

The task id field in the trace file uses the format J_C_3_4 where J stands for Julius, the second letter corresponds to the server on which the measurement has been taken (see Table above), the first number is the file that is converted (see Figure 4), and the second number is the run number.

The task id field of the MR-Leo trace looks as follows: M_1_4_78. The M stands for MR-Leo, the first number identifies the video, the second number identifies the run number and the last number is the frame number.