As discussed in the EDC's previous article, load or stress testing a ThingWorx application is very important to the application development process and comes highly recommended by PTC best practices. This article will show how to do stress testing using the ThingWorx C SDK at the Edge side. Attached to this article is a download containing a generic C SDK application and accompanying simulator software written in python. This article will discuss how to unpack everything and move it to the right location on a Linux machine (Ubuntu 16.04 was used in this tutorial and sudo privileges will be necessary). To make this a true test of the Edge software, modify the C SDK code provided or substitute in any custom code used in the Edge devices which connect to the actual application.
It is assumed that ThingWorx is already installed and configured correctly. Anaconda will be downloaded and installed as a part of this tutorial. Note that the simulator only logs at the "error" level on the SDK side, and the data log has been disabled entirely to save resources. For any questions on this tutorial, reach out to the author Desheng Xu from the EDC team (@DeShengXu).
Within ThingWorx, most things represent remote devices located at the Edge. These are pieces of physical equipment which are out in the field and which connect and transmit information to the ThingWorx Platform. Each remote device can have many properties, which can be bound to local properties. In the image below, the example property "Pressure" is bound to the local property "Pressure". The last column indicates whether the property value should be stored in a time series database when the value changes. Only "Pressure" and "TotalFlow" are stored in this way.
A good stress test will have many properties receiving updates simultaneously, so for this test, more properties will be added. An example shown here has 5 integers, 3 numbers, 2 strings, and 1 sin signal property.
curl -O https://repo.anaconda.com/archive/Anaconda3-5.2.0-Linux-x86_64.sh
conda create -n stress python=3.7
source activate stress
conda install -n stress -c anaconda psutil conda install -n stress -c anaconda requests
sudo chown -R ubuntu:ubuntu /opt/csim
sudo mv /opt/csim/csdkbuild/libtwCSdk.so.2.2.4 /usr/lib
python simulator.py --help
Set-Up Test Scenario:
Canary_Int: 10 Canary_Num: 5 Canary_Str: 5 Canary_Sin: 5 Start_Number: 1 End_Number: 200
python simulator.py ./simulator.json -i
python simulator.py ./simulator.json -k
python simulator.py ./simulator.json -c
Run Stress Test:
python simulator.py ./simulator.json -l
python simulator.py ./simulator.json --launch
Stop and Clean Up:
python simulator.py ./simulator -k
python simulator.py ./simulator -d
That concludes the tutorial on how to use the C SDK in a stress or load test of a ThingWorx application. Be sure to modify the created Thing Template (created in step 6 of the "Set-Up Test Scenario" section) with any business logic required, for instance events and alerts, to ensure a proper test of the application.
Great post! Can anyone help me locate the SteamSensor app C source code file used for this load testing activity?
You guys make the cool stuff. I have also seen the PDF with the maximum ingestion rate and there was a high correlation to the "decompile" of the infotabel into different values. Therefore CPU was the bottleneck. Is there also information available what happens and if it is possible to stream up to 150k/sec from 10-15 Kepware servers into InfluxDB servers. My assumption is, that is might be possible with a machine ~2x more CPU like the one used in the benchmark
Thanks for the feedback, Stefan!
One of the next simulations we are designing will be similar to the scenario you are describing - multiple factories connected via KepServer(s) to a cloud-deployed ThingWorx instance. I'd be happy to connect with you to talk more about the approach we are taking there.
The real balancing act in these high-speed data ingestion scenarios is two things:
User requests will put load on the platform as well, but if you can minimize the "inline data processing" that needs to run against each data point, your data ingestion rates can be much higher.
The Remote Monitoring of Assets Reference Benckmark we conducted included a decent number of expression rules being processed against each data point as they came in. If you reduced that logic - or switched to a more analytics model where you look for trends and patterns as opposed to individual data points - you would likely be able to support higher data rates on the same infrastructure.
it might be a temporary fix to enable the connection servers to ingest data that just need to be stored into the influxdb. This might be done through a direct write to the influxdb, which I think is not possible because of different architecture limits. Or we can organize it at the connection server into a better way and stream it to the main server efficiently to store it faster than today. if someday in future the active-active also can work as a load balancer this should be fine too. But I have strong concerns regarding the timeline.
While running the websocket traffic using the Python & C simulator (csim), the server reached its CPU limit when approximately 450 connections were established. The system load averages were vey high. On investigating it further, I saw that the processor run length queue and the system context switches per second were very high.
[root@BLTXLL-IDMA0228 ~]# vmstat -w -S M 10 100
procs -----------------------memory---------------------- ---swap-- -----io---- -system-- --------cpu--------
r b swpd free buff cache si so bi bo in cs us sy id wa st
805 0 0 5039 3 1477 0 0 0 2 3 4 9 11 80 0 0
429 0 0 5038 3 1477 0 0 0 0 70491 629581 42 58 0 0 0
430 0 0 5039 3 1477 0 0 0 0 70458 628943 42 58 0 0 0
427 0 0 5039 3 1477 0 0 0 0 70445 627990 42 58 0 0 0
403 0 0 5039 3 1477 0 0 0 1 70593 630295 43 57 0 0 0
422 0 0 5039 3 1477 0 0 0 1 70501 629272 43 57 0 0 0
429 0 0 5039 3 1477 0 0 0 0 70469 629137 42 58 0 0 0
446 0 0 5041 3 1477 0 0 0 0 70490 628577 42 58 0 0 0
500 0 0 5042 3 1477 0 0 0 0 70544 628094 43 57 0 0 0
453 0 0 5042 3 1477 0 0 0 0 70548 626440 44 56 0 0 0
451 0 0 5040 3 1477 0 0 0 0 70443 629529 43 57 0 0 0
433 0 0 5040 3 1477 0 0 0 0 70535 628144 41 59 0 0 0
443 0 0 5040 3 1477 0 0 0 0 70534 627692 41 59 0 0 0
[root@BLTXLL-IDMA0228 ~]# cat /proc/loadavg
431.23 425.81 422.87 557/5116 105907
It seems like by spawning a large number of processes, the system is unable to cope up with these many running and runnable processes and most of the CPU time is being utilised in context switching.
CPU - 4 core
RAM - 8 GB
OS - RHEL 7.7
Thank you for your comment. I'm the original person who developed this test approach (Draft) and appreciate comment/recommendation for further enhancement.
Your observation is correct and it's anticipated. Few quick comments:
#1) Always run simulator in dedicated VM, please don't run it with TWX server together.
#2) C-SDK may use 3-5 threads at OS level for one instance. Running 450 things on a 4 core/8G memory box will exhaust CPU for sure.
#3) Interval time also impacts CPU/Memory consumption.
If I remember correctly, 200 instances with 5 seconds internal time could be barely supported stably on 4 Core/16G memory Ubuntu 16.04 box 2 years ago.
Kindly please let me know this answers your question or not.
Thank you for the response.
Yes, your post does answer my questions. As per my findings, it is the CPU, not the memory, which is the limiting factor for running the simulation traffic. Increasing the number of CPU cores comes at a premium, and becomes difficult to justify for test machines.
Since I wanted to achieve a higher simulation traffic throughput, I modified (well actually ported it to Java) the design of the simulator from multi-process to multi-threaded utilising the same settings.json configuration. With the change, I was able to achieve 10k connected things in the Thingworx server.
Currently I have only ported the websocket traffic module. If it is of interest to you, I can share the code for your review.
You can directly use Java SDK to build simulator and it can perform way more better than C-SDK. If you mainly want to simulate websocket traffic, Java SDK should be the best choice, or the only choice, but not C-SDK.
There are several Java-SDK based simulator already in the field I believe.
I understand. But what could be the reason for Java SDK performing better than C SDK? Is it something related to C-SDK being targeted for constrained IoT devices?