1brc VS xxHash

There are some good ideas for this type of problem here: https://github.com/dannyvankooten/1brc

After you deal with parsing and hashes, basically you are IO limited so mmap helps. A reasonable guess is that even for the optimal CUDA implementation, because there is no compute to speak of other than a hashmap, the starting of kernels and transfer of data to the GPU would likely add a noticeable bottleneck and make the optimal CUDA code slower than this pure C code.

c dominates every other language again...https://github.com/dannyvankooten/1brc#submitting

You can run the bin/create-sample program from this C implementation here: https://github.com/dannyvankooten/1brc

It’s just the city names + averages from the official repository using a normal distribution to generate 1B random rows.

> GPU Hash Table?

How bad would performance have suffered if you sha256'd the lines to build the map? I'm going to guess "badly"?

Maybe something like this in CUDA: https://github.com/Cyan4973/xxHash ?

Examples of fast, simple hashes that are independent enough includes murmur, xxHash, Fowler–Noll–Vo hash function and many others

If you're only using the hash for non-cryptographic applications, there are much faster hashes: https://github.com/Cyan4973/xxHash

I found the benchmark result of AMD ryzen 5950X

But what's the catch, you're thinking ? Well, it is a bit slower than its counterparts when it comes to deserializing (and marginally faster for serializing). To achieve smaller footprint, it uses a few tricks and notably a custom hash table to deduplicate strings. This comes at a cost of course (even when featuring xxHash to speed things up), but keeps the slowdown reasonable (I think).

Non cryptographic hashes has collisions, for example, assume you having content like "abcdefg" which hashed value is "123", in case of weak hash algorithm some other content like "abcdefZ" can also have a hash "123" which basically means such hash function is failed to be unique fingerprint of particular content. BLAKE3 for example can do 6-7Gb/s which make it pretty fast and secure. If your requirement accepts collision with defined error rate, I would advise you to take a look at XXH3 if you need very snappy hash algorithm, which can run at pace or RAM access (30GB/s+), but again, run tests at particular equipment you targeting, may be AES hardware accelerated MeowHash will serve you better.