I need to scan very large JSONL files efficiently and am considering a parallel grep-style approach over line-delimited text.
Would love to hear how you would design it.
- How many grep-like ops per file?
- Is it interactive or run by another process?
- Do you know which files ahead of time?
- Do you have any control over that file creation?
- Is the JSONL append only? Is the grep running while the file is modified?
- How large is very large? 100s of MB? Few GB? 100s of GB? Whether or not it fits in memory could change the approach.
- You’re using files, plural, would parallelizing at the file level (e.g. one thread per file) be enough?
- How many files and how often is that executed?
The question is, what will be your limiting factor: CPU or disk I/O? Parallel processing doesn’t do much good if the workers have to wait on the disk to deliver more data. I’d start with an async architecture, where the program can do its processing while it is waiting on more data.
One additinal trick is to compress your files before writing them to disk, using some kind of fast lightweight compression like parallel gzip (pigz command) or lzop. When parsing them, you will have smaller disk reads but higher CPU usage, which will give speed advantage if you have server-class CPU with lots of cache.
Splitting file in equal parts and analyzing in threads each part is basically the only efficient option to utilize modern CPU architectures efficiently for your task that I can think about. Since I doubt that the data stored in your files can be quickly processed by the GPU(I assume that you have text data).
Can a file really be split efficiently? And is reading from multiple files on the same disk really faster than scanning a single file from top to bottom?
You don’t actually need to “split” anything, you just read from different offsets per thread. Mmap might be the most efficient way to do this (or at least the easiest)
Whether or not that’s going to run into hardware bottlenecks is a separate issue from designing a parallel algorithm. Idk what OP is trying to accomplish, but if their hardware is known (eg this is an internal tool meant to run in a data center), they’ll need to read up on their hardware and virtualization architecture to squeeze the most IO performance.
But if parsing is actually the bottleneck, there’s a lot you can do to optimize it in software. Simdjson would be a good place to start.
If the task to just read the data quickly without processing it(doing calculations, sorting, transformation, etc.), then yes, reading line by line is the fastest way. But the OP mentioned some processing operations on data, which may require additional time and computing power, thus it will be efficient to firstly load file into ram splitting it into chunks, give each thread a chunk to process and then combine results.
In fact, my first comment suggested that you can read file line by line and once enough lines were read in RAM, thread 1 can start processing them while thread 0 still reads new lines from hard drive. Once another chunk is ready, thread 2 can start processing it and so on.
In conclusion, it all depends on what exactly you need to do with data. Simply transferring it from HDD to RAM must be done by reading line by line. But processing of data can be split among cores of CPU to maximize the speed of computations.
