Question - io-uring, Glommio and Upmem peformance benefits for Noria?

[unibum]

Hi,

I have a question as to whether Noria's performance would benefit from:

1. io-uring ?

2. from the rust crate Glommio (a thread per core crate - https://www.datadoghq.com/blog/engineering/introducing-glommio/). I think the author wrote another blog about API performance and io_uring (https://itnext.io/modern-storage-is-plenty-fast-it-is-the-apis-that-are-bad-6a68319fbc1a)

3. As well as from Upmem (https://www.upmem.com/) which seems to have the most benefit (index searches) potential and smallest code change required to enable (index searches) it according to their site.

I understand the answer might be 'if you need that much performance then you should be able to afford to buy more cores rather than creating technical complexity' but some of the items above might only benefit incoming API calls from source and not Noria itself, so just want to understand at a simplistic level / answer if these would or wouldnt benefit Noria'S performance.

Thanks in advance!

[jonhoo]

My guess would be not that much, at least not as medium core counts (~20). You might see some wins in read latency, but on balance the current read path is quite high-performance as-is. Maybe 2x if things went really well due to fewer kernel crossings. If you go beyond 20 cores, being better about how load (and reads in particular) is spread across cores (e.g., with Glommio) could be important to keep up the scaling with more cores, but my guess is that you could get that scaling with some additional smarts to the thread pool (i.e., tokio). The work to actually switch to a different execution model would likely be quite significant, and I think on balance not worth it compared to improving tokio's performance at high core counts for example :)