Linux 6.6 To Make It Easier To Enable Partial SMT For POWER

[phoronix]

Phoronix: Linux 6.6 To Make It Easier To Enable Partial SMT For POWER

While Simultaneous Multi-Threading (SMT) on Intel/AMD x86_64 processors is limited to providing one additional thread per core, SMT on IBM POWER hardware can provide 4-way and even 8-way SMT for some processor models. With Linux 6.6 the /sys/devices/system/cpu/smt/control interface is being extended to allow greater control over managing partial SMT states in cases of the CPUs supporting more than 2-way SMT at Linux run-time...

https://www.phoronix.com/news/Linux-6.6-Partial-SMT-Control

[aviallon]

I am wondering what kind of workloads really benefits from many-way SMT.

[uid313]

Why can POWER do 4-way and 8-way SMT but x86 only 2-way SMT?

[uxmkt]

I am wondering what kind of workloads really benefits from many-way SMT.

I/O/network workloads, wasn't it? Reminds me of the SPARC T1 CPU back then.

[JEBjames]

Michael

Typo

Missing a space between words:

"forcing SMT on/offfrom user-space." should be "forcing SMT on/off from user-space."

[yump]

Does this work differently than offlining the SMT CPUs?

[-MacNuke-]

I am wondering what kind of workloads really benefits from many-way SMT.

Mixed server stuff like databases. It does not really help when you only do one kind of operation.

[-MacNuke-]

Why can POWER do 4-way and 8-way SMT but x86 only 2-way SMT?

Power CPUs have pretty much twice the amount of execution units per core. To keep them busy you need more threads with different workloads. Higher SMT counts to not help if you do not have the execution units to execute them.

On the other hand you get much more real x86-64 cores for the same amount of money and each x86-64 core is also faster on itself.

[brucethemoose]

I am wondering what kind of workloads really benefits from many-way SMT.

RAM heavy, cache miss heavy, low IPC loads like database stuff.

In these scenarios, the CPU is frequently twiddling its thumbs waiting for stuff from RAM. Higher levels of SMT give it more to do while waiting. It doesn't require a huge core either, Marvell briefly made modest 4-way SMT ARM CPUs for this purpose.