Linux Parallel CPU Start-Up Restored For AMD CPUs To Yield Faster Boot Times

[phoronix]

Phoronix: Linux Parallel CPU Start-Up Restored For AMD CPUs To Yield Faster Boot Times

The very promising work around parallel CPU bring-up to speed-up Linux kernel boot times with today's high core count servers and HEDT systems has been revised once more. Notable with the v7 patches is re-enabling support for this time-savings boot feature for AMD processors...

https://www.phoronix.com/news/Linux-Parallel-CPU-Boot-v7

[kozman]

Well how about that? A couple days ago there seemed like no interest from AMD in fixing it and then v7 patch arrives. Seems like they at least isolated the cause and can work around it. Hopefully in a v8 or later patch some smartypants will find a way to keep it enabled even with SEV-ES and X2APIC. Or maybe AMD will issue a firmware update to make it more feasible to work with those features enabled.

[blackshard]

Despite the "85%" of improvement, the real improvement in less than 1 seconds (600ms according to the patch email).
This is not to say that the improvement is worthless, but just to note that the problem was not so severe as it may be thought initially.

[stormcrow]

Despite the "85%" of improvement, the real improvement in less than 1 seconds (600ms according to the patch email).
This is not to say that the improvement is worthless, but just to note that the problem was not so severe as it may be thought initially.

It's only a problem on systems with very large core counts. Bringing up CPU cores serially on a home desktop is of no consequence. Bringing the cores up serially in a data center with thousands of cores in hundreds of packages takes hours.

ex: A single rackmount server crate might have 1024 cores cumulative over 4 CPU packages, plus a terrabyte of RAM plus storage. A single rack can have 8 crates (depending on size). That's 8192 cores per rack and say 50 racks in a row. Now you may have row after row after row of those in a 80k sq ft regional data center... 600ms adds up PDQ

[Mangix]

should be useful for my 5950X

[smitty3268]

Despite the "85%" of improvement, the real improvement in less than 1 seconds (600ms according to the patch email).
This is not to say that the improvement is worthless, but just to note that the problem was not so severe as it may be thought initially.

While that's true on your typical desktop class system, the last article mentioned one system with about 2,000 cores improving by 57 seconds with these patches, or about 25% of the total boot time. That's pretty substantial.