The thread count in these graphs is typically on the client side -- it's the number of concurrent client threads (connections) coming from the benchmark. So with more client threads than CPU cores/HT threads, the benchmark can show the results of increasing the amount of CPU contention/saturation.

On the server side, MySQL defaults to using a model of 1 thread per connection, plus various additional background threads for I/O, listening for new conns, replication, signal handling, purging old row versions, etc. Most of these are configurable, but it's not as simple as "running the MySQL server with N threads". Basically, if the benchmark is using N threads on the client side, then you can assume the server actually has N + M threads, as the server-side thread count is dynamic based on the workload.

Ah, this is the answer that makes the most sense. Thank you.

I would assume it's because not every thread may be busy at the same time - some may be waiting for I/O, or simple keep-alive threads that only need to do things once in a while.

running more threads than CPUs is useful whenever your workload is not 100% CPU-bound. a DB workload is likely IO-bound, so many threads will just be sittin' there waiting to be woken up a lot of the time

> no IO is performed

A DB that performs no IO? Can you elaborate?

Referring to this sentence, I think:

>To test CPU performance, I used a read-only in-memory sysbench OLTP benchmark, as it burns CPU cycles and no IO is performed by Percona Server.

Usually that means a read-heavy workload where things fit in RAM.

That's a literal quote from the article. The benchmark was done entirely done in-memory.

There is no such thing as no I/O.

If you're latency bound by the round trip to main memory then SMT will still have a noticeable benefit. The most important thing it can't help with is if you are compute bound.

Nobody calls "trips to memory" I/O though. Yes SMT can have a benefit but the OP was asking about the case with more threads than logical cores, so SMT is already maxed out (assuming all threads are CPU bound).

> SMT will still have a noticeable benefit

YMMV. SMT usually halves the cache size usable by each thread. If your memory access is heavily local, enabling SMT can hurt your performance.

Given it depends on the kernel version, I would imagine some kind of platform-specific tuning. Epyc is a NUMA platform, so perhaps improvements to the NUMA scheduler?

That's pure guesswork, though.

Caching effects. If you have an x-thread CPU and you split your workload into y >> x parts, it might happen that suddenly a workload fits into a faster CPU cache level (e.g. L1 instead of L2) and then you see huge speedups. I've seen things speed up manifold when I had 67 threads and added one more even though it was an 8-thread CPU. Keep that in mind when you do threading.