In the early days when hypertreading was introduced, the findings with HT on or off with database systems (mssql, postgresql) was quite clear: HT is not helpful, may even harm performance. For a long time, I had Xen Hypervisor running on AMD cpus with HT off.
We're now some some 15 years later, and I wonder how SMT has evolved in the meantime. The cluster I'm running now has Ryzen Zen1-5, with SMT enabled. Lately, on a Zen5 32-Core (i.e. 64 threads), the host load was around 45, with cpu usage of 70%, which appears sane on first sight, but the performance impact was more than noticeable (and MSSQL made it worse by taking bad plan decisions in that situation).
So, apparently the load should be not higher than actual core count, which makes me wonder if I should disable SMT rightaway again.
We're now some some 15 years later, and I wonder how SMT has evolved in the meantime. The cluster I'm running now has Ryzen Zen1-5, with SMT enabled. Lately, on a Zen5 32-Core (i.e. 64 threads), the host load was around 45, with cpu usage of 70%, which appears sane on first sight, but the performance impact was more than noticeable (and MSSQL made it worse by taking bad plan decisions in that situation).
So, apparently the load should be not higher than actual core count, which makes me wonder if I should disable SMT rightaway again.