Linux 4.15 Will Have A Scheduler Change To Benefit AMD EPYC

[starshipeleven]

Bit of noob question, I understand why the kernel will have optimizations for different platforms, so that all platforms can use the same kernel.

But from an end user perspective, isn't it a waste and inefficient to have code that won't be used.
i.e Having an AMD system but kernel will still have intel stuff??

Linux kernel has a large list of compilation options to enable/disable these features and anything else, so you aren't forced to have/use it.

Distros usually enable most x86 stuff when they compile the kernel they ship, just because the performance impact of having but not using some things is not noticeable. Gentoo is a distro that allows people to recompile everything to their own hardware's specification and remove all unused stuff from kernel or application at the compilation time.

But even in most distros a large amount of things are disabled at compile time, for example all code for supporting ARM, MIPS and others, and features that matter only for embedded devices.

[strtj]

x86 servers are typically 1-2 sockets. The 8-16 socket server domain are exclusively Unix and RISC running Unix (Solaris/AIX). With that said, if you have a crappy scheduler on 1-socket machine, it is no big difference to a good scheduler. However, if you have a bad scheduler on 8 sockets then performance will suffer a lot. Thus, as all Linux servers are 1-2 sockets, a bad scheduler is not a big problem. Unless you see Linux on 8 socket servers, you will not need a good scheduler. There is no need for follow up story until 4-8 sockets are common.

Do you have any evidence to back up the claim that Linux is not run on machines with more than two sockets? That seems like a very bold assertion and I'm not sure I believe it.

[nils_]

x86 servers are typically 1-2 sockets. The 8-16 socket server domain are exclusively Unix and RISC running Unix (Solaris/AIX).

Bollocks. Many people are migrating away from AIX these days and Solaris is all but dead. IBM is heavily pushing Linux on its POWER platform. That many hyperscale consumers don't really care for NUMA and instead opt for smaller servers in higher densities is just a side effect of the nature of their workload.

[nils_]

Bit of noob question, I understand why the kernel will have optimizations for different platforms, so that all platforms can use the same kernel.

But from an end user perspective, isn't it a waste and inefficient to have code that won't be used.
i.e Having an AMD system but kernel will still have intel stuff??

In many cases the difference is negligible. You may miss out on many optimizations but those CPU specific ones that are there mostly impact the size of the binaries.

[bridgman]

Do you have any evidence to back up the claim that Linux is not run on machines with more than two sockets? That seems like a very bold assertion and I'm not sure I believe it.

Obligatory xkcd comic:

[starshipeleven]

Do you have any evidence to back up the claim that Linux is not run on machines with more than two sockets? That seems like a very bold assertion and I'm not sure I believe it.

These things (they replaced SGI UV offering after HP bought them) can only run Linux, btw. Up to 32 sockets, with NUMA.

https://www.hpe.com/us/en/product-catalog/servers/integrity-servers/pip.hpe-integrity-mc990-x-server.1008798952.html

[MartinN]

Phoronix: Linux 4.15 Will Have A Scheduler Change To Benefit AMD EPYC

Linux 4.15 will be exciting for AMD Zen systems not only for working temperature reporting (finally) being in place for Ryzen/EPYC, but AMD EPYC CPUs should also benefit from a scheduler topology improvement...

http://www.phoronix.com/scan.php?pag...hed-Linux-4.15

AMD, like the Phoenix, has risen from its ashes!

[jromang]

Will this patch also benefit to Threadripper when configured in NUMA mode ?

[pavlerson]

Obligatory xkcd comic:

They are refering to the SGI Altix server. Which was a cluster. It only ran clustered workloads such as HPC number crunching. SGI themselves confirmed this in an interview if you care to look. The SGI UV3000 is a direct descendant to the Altix cluster.

[pavlerson]

Bollocks. Many people are migrating away from AIX these days and Solaris is all but dead. IBM is heavily pushing Linux on its POWER platform. That many hyperscale consumers don't really care for NUMA and instead opt for smaller servers in higher densities is just a side effect of the nature of their workload.

So you agree that Linux is more on "smaller servers in higher densities"? Great. There are not many large Linux servers. Why? Because they just recently arrived to the market. The large server has always belonged to Unix/Mainframes. The 10.000s of cores of Linux servers are all clusters, scale-out, and not a single large scale-up server.