How not to do cluster monitoring.

The world should have blinkenlights on its computer systems. That’s a given.

I wrote a couple of things. One was a Python program that pinged the four machines forming the cluster, and displayed a red or green light on a UnicornHD HAT to show their status. It worked very nicely. Then I wrote Python code to form part of any program that would be run in parallel on the cluster, which would send a signal saying whether each core was busy or not. It worked nicely, and I now had a row of 16 LEDs, in red or green, so I could see what was going on. It was very pretty.

Unfortunately, as it worked by sending a file by FTP every time a processor core changed between running and idle, it created a very effective Denial of Service attack on our network. Oops.

Now that I have thought about it more carefully, I shall be constructing a much better monitoring system, which will be based on sockets. I’ve been avoiding learning how to use them for far too long, anyway…

Later:

I tried at least umpteen example programs using sockets, and the connections were all rejected, and I couldn’t work out how to fix that. Suggestions, anyone?

Using a Python program to query the cluster computers took nearly six seconds to look at the 16 cores, hardly blinkenlights… A quick hack of a bash script, astonishingly, took almost as long. Back to trying to get sockets to work, then…

Working sockets tutorial!

At last, I found a socket programming example that worked, here!

I wanted to give Zan a tiny donation, but sadly his GoFundMe page seems defunct, and possibly the message I tried to send him also failed…

Sadly, I was then unable to work out how to accept multiple connections from the cluster computers.

Threading sockets programs!

There’s another set of client-server demos on GitHub, here, that I tested with Marvin and two of the Oysters, to confirm that it can do what I want. I can hoik code from those while retaining the program logic, and maybe get all four Oysters to send their status to Marvin, for him to display. I am not at all bothered that I am writing control system code for the cluster, instead of getting round to some fun applications of parallelism

Success!

I now have a GUI program that runs on Marvin, which takes a program developed on Marvin, deploys it on the four Pi’s that make up the oyster cluster, then uses mpirun to run it in parallel on 16 cores, with the results appearing on Marvin.

It’s clearly time to knock off and celebrate…

Raspberry Pi Cluster test

I’m just testing my Raspberry Pi cluster, to see if I have sorted out the setup properly this time. Finding the primes up to 10,000 with one core, and then sixteen cores, followed by using 16 cores to find primes up to 100,000 gave these results…

pi@oyster0:~ $ mpirun -hostfile myhostfile -np 1 python3 Programs/prime.py 10000
Find all primes up to: 10000
Nodes: 1
Time elapsed: 4.34 seconds
pi@oyster0:~ $ mpirun -hostfile myhostfile -np 16 python3 Programs/prime.py 10000
Find all primes up to: 10000
Nodes: 16
Time elapsed: 0.34 seconds
pi@oyster0:~ $ mpirun -hostfile myhostfile -np 16 python3 Programs/prime.py 100000
Find all primes up to: 100000
Nodes: 16
Time elapsed: 23.78 seconds

So, it is all working as it should now. Next step is to add blinkenlights on the supervising machine, Marvin, which has a UnicornHD HAT. After that, I want to get my GUI based supervisor working.

This leaves far too little time to flog stuff on eBay! I shall have to write a program to do that…

Update:

I ran it for the primes under a million, and it was disturbingly slow. I’d hope for something less than ten times as long as for a hundred thousand, but no!

pi@oyster0:~ $ mpirun -hostfile myhostfile -np 16 python3 Programs/prime.py 1000000
Find all primes up to: 1000000
Nodes: 16
Time elapsed: 2279.37 seconds

Almost ten minutes. I’m assuming things ended up swapping memory in and out, or Python doesn’t handle big integers very well. It’s not a problem, but it is one of the reasons I want blinkenlights…

At last, an app with a GUI!

For a while, I have been turning the camera on the Pi in the greenhouse on and off manually. By that I mean…

  • Connecting to the Pi using VNC
  • Opening the /var/tmp directory in the file manager
  • Creating a file called blind to switch the camera off, or
  • Deleting/renaming the blind file to switch the camera on
  • Disconnecting from the Pi

That’s clearly a huge faff, so I decided I’d finally make a GUI based program to do the job, using guizero. Other GUI libraries were either immensely complex or mysteriously impossible to install. I looked at a couple of example programs, and cobbled this together…

import paramiko
from guizero import App, Text, PushButton

def switch_camera_on():
    ssh_stdin, ssh_stdout, ssh_stderr = ssh.exec_command("sudo rm /var/tmp/blind")
    
def switch_camera_off():
    ssh_stdin, ssh_stdout, ssh_stderr = ssh.exec_command("touch /var/tmp/blind")

ssh = paramiko.SSHClient()
ssh.set_missing_host_key_policy(paramiko.AutoAddPolicy())
ssh.connect("server-name-here", username = "user-name-here", password = "password-here")

app = App(title="Greenhouse Camera Control", bg = "lightblue", width = 400, height = 200)
welcome_message = Text(app, text = "Click button to switch camera state.", size = 15)
onswitch  = PushButton(app, command = switch_camera_on,  text = "Camera on.")
offswitch = PushButton(app, command = switch_camera_off, text = "Camera off.")

app.display()

It does need you to have used ssh-keygen on your systems, so they can communicate. I can’t believe how easy simple stuff like this is to hack out!

Fear and Loathing with mpirun

My Raspberry Pi cluster, named “Oyster” because of something to do with the Walrus and the Carpenter, had been out of commission for months, so I eventually got to work and tried to set it up from scratch, as an alternative to checking everything over and over again, and failing to find anything wrong. Oyster used to look like this, but the Pi 3s in Lego compatible cases ran too hot.

At first, I attempted to use my only Pi 4, “Marvin”, as the control machine for the four Pi 3s in Oyster, but I got something wrong in the setup, and it didn’t work. I had been thinking 20 cores would obviously be better than 16, but suspected the two different user names in use might be causing the problem. It probably wasn’t, as I now think the ssh communication for the cluster is done anonymously. Possibly. Anyway, I changed my mind about 20 cores, when I thought about the other tasks Marvin runs, and how I didn’t want them slowed down. Sixteen will do. Unless I get some more Pi 3s and add them to the cluster…

Anyway, I went through all the setup described in Ashwin Pajankar’s e-book* about Raspberry Pi “supercomputers”, again. Twice. I found some online guides, and checked the setup with those, too. I could run programs on the four cores of oyster0, but not on the other three Pis. Eventually, I spotted that I had forgotten to create the file “known-hosts” on the controlling Pi. The message passing software, python3-mpi4py, would probably have told me this, if I had run it in verbose mode, but I didn’t. Still, it now works, and I have a sixteen core Pi cluster running.

I have added a program, running on Marvin, that lights up LEDs to show the state of the Pis in the cluster, and intend to add further blinkenlights to show the activity of each of the sixteen cores.

# Program to monitor status of Oyster cluster, displaying up/down indication.
#import os
import time
import subprocess as sp

machines = ["oyster0", "oyster1", "oyster2", "oyster3"]

while True:
    for i in range(len(machines)):
        machine = machines[i]
        state  = sp.call(['ping', '-c', '1', machine], stdout=sp.DEVNULL)
        
        if state == 0:
            colour = "0 200 0" # Green means UP
        else:
            colour = "200 0 0" # Red means DOWN 
            
        row = "12 "
        if machine == "oyster0":
            column = "0 "
        elif machine == "oyster1":
            column = "4 "
        elif machine == "oyster2":
            column = "8 "
        elif machine == "oyster3":
            column = "12 "
    
        message = "set_pixel " + row + column + colour
        
        fp = open("/home/chris/ftp/files/blinkenlights" + machine, 'w')
        fp.write(message)
        fp.close()
        
        time.sleep(1)

This code sends a text file to my Unicorn HD server program, which enables more than one program to write to its 256 LEDs, without messing up each other’s displays.

My plans for further development include improving my set of framework code for running parallel programs on the cluster, and more (or possibly less) importantly, to send messages to Marvin for the blinkenlights.

[* I haven’t included a link to AP’s e-book, because it’s easy to find online, and he is charging far too much for it, while others are giving us the same information free.]

Fun with rpi-connect

To begin at the beginning…

It all started when I read a news item saying RealVNC was going to change its terms and conditions unilaterally, so that home users would be limited to connecting to three computers. I was using it to view and control anything up to fifteen computers.

Panic set in. I began to research other free VNC implementations. Then it was pointed out to me that it was only remote access over the internet that would be affected. I hadn’t been doing that at all. It seemed all the Pi’s on my network would be unaffected. To ensure things would not get changed, I set RealVNC to NOT update automatically on all of them.

During my researches, I heard that a Pi running Wayland instead of X-Windows could be remotely accessed just like VNC, using rpi-connect, and decided that would be interesting to do, anyway. So I needed to update a machine to run the Bookworm version of RaspberryPi OS, which does use Wayland. I have only one Pi 4 at present, and did the thing they tell you never to do – I attempted an in-place upgrade. It ever so nearly worked. But it wouldn’t run rpi-connect. So, I pulled the Pi to bits, took its SSD off the Waveshare adapter it uses, and burned a fresh copy of Bookworm. Once it was all re-assembled, the Pi booted up normally, and rpi-connect was usable. Well, there was the small matter of going online and associating the Pi, whose name is Marvin, with my RaspberryPi ID. Which I have.

Here is a screenshot of my Linux Mint box, with VNC Connect at the top left, a VNC session into the street camera at the bottom left, another remote session on a Pi, at the middle left, and a remote session on Marvin…

Ideas for the next step…

Firstly, it seemed like a good idea to verify that I did indeed have remote access. I turned off wi-fi on my phone, and used Chrome to access the Raspberry Pi sign in page, which was fine. And on connecting to Marvin, I got this…

So, remote access definitely works. I don’t think I will be using my phone for the job, but one of my tablets, or even the Chromebook should work just fine.

The next thing was to write something to enable Marvin to monitor WeatherPi, which has an occasional problem with its greenhouse temperature sensor, that results in the weather monitoring and upload program crashing. I used the Python Paramiko library, and made a program (cobbled together from earlier versions) that checks the weather station every five minutes. If the machine itself is down, it sends a message to my phone, using Pushover. If the machine is up, it checks whether the weather program is running, and attempts to restart it otherwise. I added starting this program to the other programs Marvin runs at startup, which is done by this bash script –

#!/bin/bash
lxterminal --title "UnicornHD" -e 'python3 Programs/unicorn_server.py && read x' &
sleep 2
lxterminal --title "Fan HAT" -e 'python3 Programs/fan65.py && read x' &
sleep 2
lxterminal --title "Weather Station" -e 'python3 Programs/WeatherMonitor.py && read x' &
sleep 10

That was unnecessary!

Somewhere online, while I was trying to get the remote monitoring to work, I found a way to make a python program keep going. Using a bash shell script to run the program like this means that if the program does stop, it will immediately be run again, until the terminal it was run in is closed.

while : ; do
    now=$(date)
    echo $now >> restart.log
    python3 /home/pi/program-to-run.py
    sleep 10
done

This does not work with programs using Pi cameras, when the camera crashes. I wonder if there is a camera resetting utility available, as rebooting is not possible inside this script! I wonder if libcamera can do a reset of the camera? Must have a look…

Raspberry Pi 7 Segment display

Introduction

One of my #RaspberryPi Zeros is called PiClock, and has an 8 digit seven segment LED display. The program it runs displays the time, and sends it to two other Pis, that display it on Unicorn HD HATs. Between midnight and 8 am, it flashes the message “SLEEP” every five minutes, as well. The software library that it uses can display numbers, and most upper and lower case letters, but not all of them. I rather liked the idea of animating sequences of single segments on it, because, well you know, blinkenlights. I had a look at the software library, “7seg.py”, to see if I could get it to do that.

It turns out that the library uses a Python dictionary to look up the byte to send to the display for each of the characters it can display. Decoding the hexadecimal bytes took a few minutes, working from the code for the digits from 1 to 5.

The first bit is always a 0. The remaining seven are the seven segments, in the order abcdefg, which are laid out like this…

So, the codes for illuminating single segments are as follows…

Now to amend the library! I needed some typeable characters to put in the dictionary, ready to be used in strings in the python code. For no obvious reason, I chose a selection of brackets and the tilde character, and amended the library file. The selection of brackets didn’t work!

After trying characters until they did work, I ended up with #][£<$~ as the symbols for the segments abcdefg.

I’m only showing the amended part of the file, where the pattern to send to the display is looked up. The arrangement of the brackets and tilde for the segments is as follows…

Now I’m ready to program PiClock to do silly animations, which will be fun, and a lot easier than using the WordPress editor. Note to self: See if you can find a WYSIWYG editor for WordPress.