SUPERCILEX

In my benchmarks[1], the average processing time for an element is 250ns with 4 producers and 4 consumers contending heavily. That's terrible! Even if your numbers are correct, 100ns is a bit faster than two round trips to RAM while 33ns is about three round trips to L3 and ~100x slower than spamming a core with add operations. That's slow.

[1]: https://github.com/SUPERCILEX/lockness/blob/master/bags/benc... $ cargo bench -- 8_threads/std_mpmc

In the example with multiple counters, in real life each counter could shout out a number and have people approach their respective counters in parallel. But this is not how lockless queues work. Instead, the person at the head of the queue holds a baton and when multiple numbers are called, everybody waiting in the queue goes up to the counter of the person holding the baton. Once that head-of-the-queue has made it to the counter, they give the baton to the person behind them who then drags everybody along to their counter. And so on.

The article was arguing for a lockless channel implementation akin to your interpretation of a queue with parallel access to the counters.

The point was to be able to divide the day into nicely sized bites for getting stuff done. I chose 10 minutes in an hour for that reason: you can get a small task done in one "minute."

Mostly about performance and project internals.

$ hyperfine --warmup 3 -N "./test /dev/shm 8 zip" "./test /dev/shm 8 chain" Benchmark 1: ./test /dev/shm 8 zip Time (mean ± σ): 118.5 ms ± 11.6 ms [User: 92.9 ms, System: 726.6 ms] Range (min … max): 103.6 ms … 143.4 ms 23 runs

Benchmark 2: ./test /dev/shm 8 chain Time (mean ± σ): 235.7 ms ± 11.0 ms [User: 116.4 ms, System: 1537.7 ms] Range (min … max): 220.1 ms … 258.3 ms 13 runs

Summary './test /dev/shm 8 zip' ran 1.99 ± 0.22 times faster than './test /dev/shm 8 chain'

Note that I'm out of my depth here, so this is all speculation. Until we hit hardware limitations (which will be PCIe 6 if I had to guess), I'm pretty sure those are the same thing. One read/write iop = 4KiB. If your PCIe bandwidth is limited to N GB/s, then there are only so many iops you can physically send/receive to/from the SSD, regardless of how many iops the SSD could be capable of processing. So currently we're bottlenecked by PCIe, but I doubt that will continue to be the case.

> again, you're handwaving on what "interfere" means

It depends on how the file system is implemented, but my guess would be that a lock on the inode or block cache entry is acquired.

> do "inode mappings" represent resources with no shared resource constraints?

Those are the contents of the directory. The problem is not reading them, but changing them.

Also no need to theorize: run the benchmark I linked for yourself. It clearly shows a massive advantage to having each thread work with its own directory.

Plus accompanying benchmark: https://alexsaveau.dev/blog/projects/performance/files/fuc/f...

---

> file operations are (almost always) io bound

This is a common misconception. It was presumably true a decade ago, but PCIe is getting exponentially faster every 3 years: https://arstechnica.com/gadgets/2022/06/months-after-finaliz...

The NVMe protocol has extremely deep queues [1] and leaving them empty means leaving performance on the table. I think you'll find it surprisingly difficult to saturate the PCIe bus with just one core: PCIe 7 will support 512GB/s. Assuming a single core can produce 64 bytes (an entire cache line!) per cycle running at 5GHz, you're still only at 320/512=62.5% saturation. This napkin math is a little BS, but my point is that individual cores are quickly going to be outpaced by bandwidth availability.

> and totally unclear how directories represent an "interference" boundary

To add a bit more color here, it depends on how your file system is implemented. I belive Windows stores every file's metadata in a global database, so segmenting operations by directory yields no benefits. On the other hand, Unix FSs tend to store file_name to inode mappings per directory, so creating a new mapping in one directory doesn't interfere with another directory.

[1]: https://en.wikipedia.org/wiki/NVM_Express#Comparison_with_AH...

"The macOS/Windows implementations are currently equivalent to the *_rayon implementations shown in the benchmarks."

Rayon is pretty good, but clearly suboptimal as evidenced by the benchmarks.

FTZZ is designed (and optimized) primarily for Linux, but also works on macOS and Windows.

FTZZ is insanely fast and will saturate your machine when creating empty files — I can reach 800K files/s on my laptop. I'm considering the project complete, but it still has a few flaws: - Memory usage is not fully bounded, meaning you could theoretically run out if you try to generate a tree billions of directories wide. - The error between the target number of files and the actual number of generated files grows with the maximum depth of the tree. I'm using a LogNormal distribution, but probably incorrectly. - When generating an exact number of files or bytes, their distribution can get messed up because I haven't looked into how to properly distribute the remaining files/bytes.

If you want a more in-depth dive into the technical and performance details, I wrote up a blog post about them: https://alexsaveau.dev/blog/projects/performance/files/ftzz/.... If you think this project is cool, please consider upvoting my Stack Overflow answer: https://stackoverflow.com/a/69714764/4548500.