Joseph Bellahcen

This post originally appeared on my website.

Today I’m retiring my beloved Raspberry Pi homelab, hostname pigeon. I bought it back in 2023 after the stock of Pis finally recovered from the COVID-19 pandemic. It’s always functioned as a homelab, perfectly quiet and low-power, tucked away under a shelf on my desk, crunching away at the bits and bytes of my digital life.

It hosts a handful of Docker-based applications including a services directory (Madness), an RSS feed aggregator (FreshRSS), a YouTube archiver (Pinchflat), a file server (NextCloud), and an Obsidian notes server (CouchDB). These services help me exercise sovereignty over my data and preserve my privacy against the prying eyes of corporate monoliths. They empower me to be aware of and choose where my news comes from, to maintain copies of my favorite videos even if the platform deems them unsavory, and to own my digital files rather than renting space for them on someone else’s computer.

However, there are still some areas where the Pi’s modest specs limited its utility. I tried to spin up a backup solution for my photo library (Immich) and personal movie collection (Plex), but media processing pipelines proved too demanding. Instead, I’ve been paying for iCloud storage for my photos and using the Home Sharing feature on my Mac mini to stream to my Apple TV. And despite my reservations about the rising prevalence of large language models (LLMs) like ChatGPT, they are becoming increasingly capable and I’d like to give them a fair chance. Thus, I need more powerful hardware to evaluate the viability of hosting open-weight models like GPT-OSS and Qwen, so I can leverage the power of this emergent technology without sacrificing my privacy.

To bridge the performance gap and expand the range of services offered by my lab, I’ll be migrating over to the Albatross. This PC build has a beefier Intel i9-12100K CPU, NVIDIA RTX 3060 GPU, and 32 GB of RAM. I’ve migrated my NextCloud instance, and I plan to extend it with a Google Docs-like web editing experience via ONLYOFFICE. I’ve also spun up an LLM chatbot application with OpenWebUI that can run larger models (30B parameters) with good throughput (up to 50 tokens/sec). I’m going to give Immich another shot and hopefully cancel that iCloud subscription. Although I’ll likely continue using Home Sharing on my Mac mini to stream movies, since the files remain on my device.

This post originally appeared on my website.

why are checksums useful for fpga development?

Checksum functions are useful for error detection and data integrity. With respect to FPGA architecture development, a checksum can ensure the system is able to retain its state. For example, an FPGA programmed to behave like an AND gate should not start to act like an OR gate due fluctuations in voltage or temperature. Such problems may be caught by computing the checksum of the bitstream before configuring the FPGA, operating the device in the lab, and then extracting the bitstream and re-computing the checksum to ensure it has not changed.

The last two FPGAs that I worked on-with Indiana University's SAIL-IN Lab and the QuickLogic Corporation-used a scan chain configuration interface which lent itself well to bitstream verification via checksum. The scan chain acts like a serial shift register, stringing together all the configuration chain flip-flops in the FPGA fabric. As the bitstream flows through the head of the scan chain, one bit at a time, its checksum may be computed; and later re-computed as it flows out of the tail. For context, imagine a small, thirty-two flop scan chain and the corresponding bitstream, 0xdeadbeef. On the way in, the Fletcher's (32-bit) checksum will be 0xf13b. If the post-configuration checksum does not match, the engineers would know that there is an issue with the FPGA architecture.

Fletcher's checksum also produces check bytes (or a "tag"), which can be appended to the end of the original data, such that the new checksum of the data and the check bytes is zero. For the example bitstream 0xdeadbeef, the resulting check bytes are 0xd2f1. And the checksum of 0xdeadbeefd2f1 is 0x0.

fletcher's checksum in systemverilog

Based mostly on the Wikipedia entry for Fletcher's checksum, I produced a parameterized, three-state module which computes the checksum and check bytes of a data stream. The SystemVerilog closesly follows the below psuedo-code. For the Fletcher-32 checksum, data arrives as 16-bit half-words.

Eventually, I'd like to wrap it in a SPI interface and submit it to TinyTapeout, but for now here's the bare RTL. The most up-to-date code is also available on GitHub:

This post originally appeared on my website

I was born shortly after the peak of the dot com bubble, after which point "google" became a verb and "facebook" was no longer something collecting dust in the attic. During this period, it felt as though the noisy construction of the modern Web was occurring in the next city over and not yet in the quiet towns where most of us resided. Recently, following the COVID-19 pandemic, many of us experience (social) media fatigue, and it lulls us back to when the Internet was more peaceful. To a time when most communication happened over email or XMPP, when we "followed" people by subscribing to their RSS feeds, and when a personal homepage was an extension of self.

This nostalgia for a simpler and more personal Internet is a driving factor behind what is known as the "small web" or "indie web." The small web is a loosely defined philosophy which promotes decentralization, user agency, and community. It does not intend to displace the tech giants of today, but instead it empowers people to host their own content and collaborate through alternative platforms. Many concepts and technologies fall under this umbrella: personal blogs, community forums, webrings, federation, etc. The ideology bears an intrinsic link to open-source software; and its echoes are audible in the alternative web protocol Gemini, federated platforms such as Mastodon and Lemmy, public-access UNIX boxes (also referred to as tilde communities), and more.

The school bells are ringing! I am so looking forward to this year, especially now that things are starting to open up due to the availability of vaccines. I will be continuing in my role as an RA, but I now manage a floor comprised entirely of Computer Science and Engineering majors. The opportunity to tailor the residential curriculum specifically to topics in computing is going to be an amazing way to bring my own passions into the community. This year, I will also begin taking classes toward my master's degree in Computer Engineering, meaning that I will finish undergrad with half of my master's already out of the way!

Pictured in this post are some artifacts that I created this summer as a research assistant in a computing lab on campus. Our team is developing field programmable gate arrays (FPGAs) for defense applications. If you don't know, an FPGA is an integrated circuit that can be programmed to behave like any other circuit using a hardware description language like Verilog. They can be found in a host of devices you interact with every day (TVs, radios, cameras) as well as military-grade equipment like what I am working on. They are widely used due to their ease of programming, which allows for rapid and cost-effective development, as well as the ability to patch bugs without having to re-manufacture the circuitry.

In the lab, I maintain a custom toolchain that converts Verilog code into a bitstream that can be loaded onto our FPGA. At the moment, this is all happening in the context of a simulation, but by the end of the fall our little black box should come home from the factory and I can hold it in my hands! At that point, I will begin designing some circuit boards that make use of the device before it eventually makes its way onto a military drone or something of the sort.

Okay, I am having way too much fun with this project… 

As a part of my cloud computing course, I have to train and deploy a machine learning solution. I chose to create a password strength classifier, which consists of two major parts. 

The first is an input vectorizer, which accepts text input and converts it to an array of floats that describe how the characters in your password relate to one another and to other passwords. The second part is the machine learning model, which is a multi-layer perception classifier. The model was fed over 500k passwords and their associated strength rating, which it used to fine tune the 500 neurons in each of its hidden layers. The training process took almost four hours, but as you can see by the F1 score and confusion matrix, it was worth the wait! 

This is my first time developing a machine learning model without a tutorial or textbook, and honestly the experience has been a blast so far. Even though my system is quite simple, building it reinforced a lot of the skills that I have been learning throughout this academic year. It also taught me how to vectorize text for deep learning, which is something I hope to apply to develop chatbots and NLP applications in the future. 

Handwriting update!

Thank you to @mourza​ for the question on my previous post. Here, I will try my best to explain how reverse engineering works for Mourza and anyone else who might be interested. Full disclosure: I am NOT an expert on the matter, just an enthusiastic student who learned about this last semester.

But First, A Review of Assembly

As you may know, computers can only think in binary. That is, your CPU can only understand 1 or 0, high or low, on or off. Even though anyone can write software using a high-level language like C, Java, or Python, at some point it will need to be compiled down to those 1s and 0s that can talk directly to the CPU.

Take, for example, the following C-like statement.

int x = 49;

When this statement is compiled into a program, it will be converted to assembly language, which will allow your operating system to bind binary instructions directly to the CPU. Each brand of processor has a unique assembly language, but here is an example of what the code might look like after compilation.

movl $49, -8(%rbp)

While this may not look like anything meaningful at first glance, we can go ahead and interpret it. In essence, the line is going to move the numerical value of 49 into the register known as the base pointer. If you want to learn more about the assembly instructions for 64-bit Intel processors (the kind you probably have in your computer), check out this reference sheet.

Reading Assembly with Radare

At this point, you might be wondering how you can see this assembly code. If you try to open up a .EXE file in a text editor, you will not see a statement like the above. Instead, you might find a bunch of garbled nonsense. What you would be looking at are the binary instructions that will go to your CPU, but your text editor is trying to interpret them as text.

If you want to see the assembly, you have to use a decompiler such as radare2. I won’t go into the nitty gritty of how to install or use it, since you can read about that on their website. That being said, if you download the reference sheet PDF I linked above and go to my previous post, you can probably read the screenshot and figure out the password to diffuse the “bomb” program. I challenge you to comment on that post with the password!

I have to break it to you, though. Most programs are not quite as easy to decompile and read. The binary bomb is an assignment designed to help students like myself practice reading assembly. Most programs you use every day are insanely complex, and it just wouldn't be possible to navigate their instructions. But for smaller executables, especially malware, it can be possible to peek inside and figure out how they are causing damage.

One of the nice things about being stuck in quarantine during the holidays was that I finally had the time to play around with circuits. When I was a kid, I remember playing with Snap Circuits and building noise-makers and AM radios, and this really feels like the big boy version.

After countless hours and way too many microwavable meals to get me through them, I completed the second and final semester of physics required for my computer engineering degree. While it wasn’t easy, the intense course load and complicated material helped me strengthen my studying and time-management skills. Especially with courses being taught online, being able to motivate myself to engage with the material became very important, and I think I have gotten good at it.

Happy 2021, everyone!

This academic year, I have been working as a Resident Assistant on campus. The job is a lot different than I was expecting due to the pandemic, plus I am stationed at the university’s “party dorm,” but I still enjoy it. I have thirty residents who are all very kind and academically driven, and we all get along.

Hello, everybody!

I am finished with my Linear Algebra course, wooo! The class was tons of fun, and I gained some useful insights that I will be applying to a programming project that I have started surrounding Linear Predictive Coding! Plus, I got to test-drive more online classes before that becomes everyone’s reality starting in August...

Since I am incapable of handling free time between now and then, I have started to take a Physics course offered for free by MIT, and I am hoping to pre-learn some of the material before I take my university’s electromagnetism class next semester. Today, I learned about electric fields and dipoles. Interesting stuff!

I am also experimenting with some new note-taking styles, which I hope will make it easy to cross-reference these notes later during the school year. If you have any creative tips, definitely share them with me so I can give them a try!