I was diving into PoW (Proof-of-Work) once again after nostr:nprofile1qy88wumn8ghj7mn0wvhxcmmv9uq3wamnwvaz7tmjv4kxz7fwdehhxarj9e3xzmny9uqzqj8a67jths8euy33v5yu6me6ngua5v3y3qq3dswuqh2pejmtls6datagmu rekindled my interest with his PoW Draw project. It was a fun little trifle, but it shifted my focus just the right way at the right time.
Because then, on Friday, came the [Oval Office Travesty](nostr:nevent1qvzqqqqqqypzpmym6ar92346qc04ml08z6j0yrelylkv9r9ysurhte0g2003r2wsqy2hwumn8ghj7un9d3shjtnyv9kh2uewd9hj7qghwaehxw309aex2mrp0yhxummnw3ezucnpdejz7qpqqqqqqqrg6vz7m9z8ufagn4z3ks0meqw4nyh4gfxvksfhne99egzsd3g3w9). Once I got over the initial shock, I decided I couldn't just curse and lament; I needed to do something bigger, something symbolic, something expressive. So that's exactly what I didโbreaking nostr:nprofile1qy88wumn8ghj7mn0wvhxcmmv9uq32amnwvaz7tmjv4kxz7fwv3sk6atn9e5k7tcqyqewrqnkx4zsaweutf739s0cu7et29zrntqs5elw70vlm8zudr3y2t9v7jg's record which he held for almost 2 and half years.
Here is a note with PoW 45, the highest PoW known to Nostr (as of now).
nostr:nevent1qvzqqqqqqypzpmym6ar92346qc04ml08z6j0yrelylkv9r9ysurhte0g2003r2wsqy88wumn8ghj7mn0wvhxcmmv9uqsuamnwvaz7tmwdaejumr0dshsqgqqqqqqqqqy8t8awr5c8z4yfp4cr8v7spp8psncv8twlh083flcr582fyu9
## How Did I Pull It Off?
In theory, quite simple: Create note, run PoW mining script & wait.
Thanks to PoW Draw, I already had mining software at hand: nostr:nprofile1qy88wumn8ghj7mn0wvhxcmmv9uq32amnwvaz7tmjv4kxz7fwv3sk6atn9e5k7tcqyqvqc5tlvn6etv09f0fvuauves49dvgnukjtzsndfv9y8yyrqyxmz7dty6z's [*notemine_hw*](https://github.com/plebemineira/notemine_hw), but when you know that there is a 1 in 2^45 chance that the next hash will be the correct one you want to increase the odds a bit. So on Monday evening, I started my Note Mining operation on an old 40 thread machine called Workhorse.
### Issues Along the Way
I was immediately surprised that Workhorse (2ร Intel Xeon Silver 4114) produced only about 3Mh/s. A laptop (Intel Core i7-1185G7) with Windows and all the bloat did 5Mh/s. That was strange.
Another hurdle was that *notemine_hw* does not refresh the `created_at` field. With just a few Mh/s of power I was potentially looking at weeks of computation, by then the note would be quite stale. So I created systemd service leveraging the `RuntimeMaxSec` option to periodically restart every 3600 seconds assuring that the Note would be max 1 hour old at the time of publishing.
Luckily PoW is that kind of problem where every hash attempt is an independent event, so the chance of success is the same whether you do it in small increments or one uninterrupted stretch. So by restarting the mining process I was only losing a few mere seconds every hour due to the overhead.
Once the note staleness issue was resolved, I looked at the 40 workers on Workhorse vs. 7 workers on the laptop and start messing around with running one instance with 40 workers and running 40 instances with 1 worker and found out, that the workers are not bound to a CPU thread and are jumping between the CPUs like rabbits high on Colombian carrots.
The solution? Running multiple instances with one worker each as a service locked to its own CPU core using systemd's `CPUAffinity` option.
```
$aida@workhorse:systemd/system $ sudo cat notemine@.service
[Unit]
Description=Notemine HW Publish (restarts hourly)
[Service]
Type=simple
CPUAffinity=%i
# The command to run:
ExecStart=/home/aida/.cargo/bin/notemine_hw publish --n-workers 1 --difficulty 45 --event-json /home/aida/note.json --relay-url 'wss://wot.shaving.kiwi' --nsec nsec0123456789abcdef
# Let the process run for 1 hour (3600 seconds), then systemd will stop it:
RuntimeMaxSec=3600
TimeoutStopSec=1
# Tells systemd to restart the service automatically after it stops:
Restart=always
RestartSec=1
# run as a non-root user:
User=aida
Group=aida
[Install]
WantedBy=multi-user.target
```
Then I added a starting service to spawn an instance for each CPU thread.
```
$aida@workhorse:systemd/system $ sudo cat notemine_start.service
[Unit]
Description=Start all services in sequence with 3-second intervals
[Service]
Type=oneshot
ExecStart=/usr/bin/zsh /home/aida/notemine_start.sh
RemainAfterExit=yes
[Install]
WantedBy=multi-user.target
```
Here is the startup script (I know, loops existโbut Ctrl+C/Ctrl+V is so old-school):
```
aida@workhorse:~ $ cat notemine_start.sh
/usr/bin/systemctl start notemine@0.service
/usr/bin/sleep 3
/usr/bin/systemctl start notemine@1.service
/usr/bin/sleep 3
/usr/bin/systemctl start notemine@2.service
/usr/bin/sleep 3
/usr/bin/systemctl start notemine@3.service
/usr/bin/sleep 3
...
...
...
/usr/bin/systemctl start notemine@38.service
```
The sleep there is critical to make sure that the `created_at`timestamps are different, preventing redundant hashing.
**This adjustment made Workhorse the strongest machine in my fleet with 10+Mh/s.**
## The Luck Aspect
From Monday evening, I started adding all machines at my disposal into the fleet and by Wednesday evening I was crunching hashes on about 130 CPU threads (a lot of them were quite antique) and at the peak was just little shy of 40Mh/s. To compensate for the slow start with the few above-mentioned hiccups and the fact that I had to use my desktop to do other things from time to time, I counted with the conservative estimate of 30Mh/s when I was doing all the probability calculations.
Based on the type of task that PoW mining is, the outcome is not predictible. You are only looking at what is the chance that the outcome of every single independent event will be consecutively non-favourable and then subtracting it from 1 to get the chance of that single favourable event you want. I really had to brush up on my combinatorics and discrete mathematics to make sure I have at least an elementary understanding of what is going on. Also, because we are not just throwing a dice 5 times, but are operating with big numbers, approximation was necessary. Luckily, the formula is available and quite simple in the end.
Two weeks to exhauste all the possible tries still doesn't guarantee anything, actually there is a slighlty less than 2 in 3 chance that you will have a result after all that time. So the fact that I was able to hit the right hash in less than 3 days was good luck. Not insane lottery winning luck, but good luck; slighlty lower than 1 in 5.
## Do you want to beat me?
Go ahead! All the pitfalls are described above and until there is a GPU-based PoW Mining available, we are all on pretty even ground.
## Do you hate the note?
In that case, feel free to enjoy this accompanying image:
