Does having the worker pool hold as many threads as cores work well alongside the async pool? It is basically oversubscribed by design.
I built a system once which had (this is Rust) a Rayon worker thread pool of 4 threads and a Tokio async pool of 2 (multithreaded runtime). On a system of 6 vCPU. This ended up working fine. Tokio was not starved so handled network requests at low latency.
One difference is DuckDB is a pure network client. If one of its async threads is starved it is not the end of the world (e.g. k8s does not kill your pod for failure of replying to health checks).
Network IO is heavily NIC queue bound, if your NIC only has one queue it just makes it slower to do any threading workload against it.
On my ryzen 9 it needs around 8 cores to do the same work in a threaded io loop than you can do single threaded. And the mechanism doesnt matter, you could share an fd, use SO_REUSEPORT or just share memory between threads.
Just doing the sharing makes everything extremely slow. One context switch becomes more expensive than just doing it single threaded.
I've run quite a few benchmarks on that as well, on a few different machines, and oversubscribing ASYNC threads demonstrated very little performance downside. In the end, the memory governor also keeps these threads "in check" while still allowing full utilization when possible.
There is still something to gain from tuning it further (as you can see in the async I/O tuned benchmark), but having that network saturation by default is still a work in progress.
As long as you're scheduled by the kernel and not something like Kubernetes with a CPU limit, you can generally oversubscribe I/O threads without much of a problem. They're mostly parked waiting for syscalls anyway. Heck, even if they're mostly doing CPU work, the scheduler generally deals with it pretty gracefully.
Using 512gb of ram for a 22gb remote file does feel a bit weird for a benchmark but maybe they couldn’t get a large number of cores without lots of memory?
The main reason I decided to use a beefier machine is that it gives me flexibility when benchmarking, without the need to set up different environments. The CSV data, for example, is >80 GB. We can also “scale down the machine” for experiments where we want to stress-test lower-memory scenarios or use fewer threads by configuring DuckDB’s settings (e.g., SET memory_limit = '10GB'; or SET threads TO 1;).
Most cloud providers start with a 2:1 ratio of memory in GiB to CPU cores and go up from there. Databases also are the most common workload for large-memory systems because they benefit so much from large buffer caches.
Do the CSV files allow quoted newlines? If yes, what's the trick to avoid checking the whole file too find out whether a newline is quoted or a record separator when reading it from the middle in an async thread?
DuckDB uses a speculative parallel CSV parsing technique. The basic idea is that the parser speculates about the state the CSV parser is in at a random byte (e.g., whether it is inside a quoted field) and tries to figure out where the next row starts based on that.There are validation steps during finalization as well, to ensure the parser did not got anything wrong in its speculation.
I can't say it's 'deep' in any way, e.g. what type of queue is that? Another part - utmost importance to keep async threads fully busy - that doesn't address what kind of disk subsystem is and how it deals with random reads.
I built a system once which had (this is Rust) a Rayon worker thread pool of 4 threads and a Tokio async pool of 2 (multithreaded runtime). On a system of 6 vCPU. This ended up working fine. Tokio was not starved so handled network requests at low latency.
One difference is DuckDB is a pure network client. If one of its async threads is starved it is not the end of the world (e.g. k8s does not kill your pod for failure of replying to health checks).
On my ryzen 9 it needs around 8 cores to do the same work in a threaded io loop than you can do single threaded. And the mechanism doesnt matter, you could share an fd, use SO_REUSEPORT or just share memory between threads.
Just doing the sharing makes everything extremely slow. One context switch becomes more expensive than just doing it single threaded.
There is still something to gain from tuning it further (as you can see in the async I/O tuned benchmark), but having that network saturation by default is still a work in progress.
(Disclaimer: I'm the author of the blogpost)
(Disclaimer: I’m the author of the blog post.)
I've never gotten around to writing a blog post about it, but I go quite in-depth on the technique in this presentation: https://www.youtube.com/watch?v=YrqSp8m7fmk
(Disclaimer: I'm the author of the blog post and also the developer who implemented the entire CSV parser in DuckDB.)
Ducks all the way down!