Loongson Preparing LoongArch KVM Virtualization For Linux 6.7

[phoronix]

Phoronix: Loongson Preparing LoongArch KVM Virtualization For Linux 6.7

The past several Linux kernel cycles has seen Loongson engineers working on implementing more kernel features for LoongArch, the MIPS-derived and RISC-V-inspired architecture out of China for domestic PCs. The performance of LoongArch CPUs so far still aren't competitive to x86_64 or Arm hardware but at least slowly are becoming more practical with more features being wired up. The latest milestone is the Kernel-based Virtual Machine (KVM) virtualization support appearing ready ahead of Linux 6.7...

https://www.phoronix.com/news/LoongArch-KVM-For-Linux-6.7

[alpha_one_x86]

I have undertand LoongArch have crossarch instruction, this will allow KVM with another arch (x86), near native speed?

[xen0n]

I have undertand LoongArch have crossarch instruction, this will allow KVM with another arch (x86), near native speed?

The binary-translation assists are meant to be utilized by user-space JITs like QEMU, and not directly usable for x86/ARM emulation. Right now QEMU is not using them but I have plans to add support for some; the others are a bit too intrusive to the current QEMU TCG architecture (Loongson themselves use a heavily modified QEMU fork) so perhaps a brand-new emulator would have to be written to realize all the potential performance.

The work is certainly going to take a few years but the Loongson Corporation is only ~1000 people (of them only ~600 are devs), and the open-source community is small right now, so perhaps one can't demand more. Their next-generation models brings some real performance improvement though (I have one at hand), let's see how it goes after a few months...

[JEBjames]

Michael

Typo/Grammar

"Given the current levels of LoongArch CPU performance, the KVM virtualization support isn't likely too practical today but will become more important as the performance improved and if Loongson is to make a competitive splash in the server landscape for cloud/virtualization."

should be "improves"

[stormcrow]

The binary-translation assists are meant to be utilized by user-space JITs like QEMU, and not directly usable for x86/ARM emulation. Right now QEMU is not using them but I have plans to add support for some; the others are a bit too intrusive to the current QEMU TCG architecture (Loongson themselves use a heavily modified QEMU fork) so perhaps a brand-new emulator would have to be written to realize all the potential performance.

The work is certainly going to take a few years but the Loongson Corporation is only ~1000 people (of them only ~600 are devs), and the open-source community is small right now, so perhaps one can't demand more. Their next-generation models brings some real performance improvement though (I have one at hand), let's see how it goes after a few months...

This is all true. However, Loongson Corp. is pretty much funded by the PRC state (through at one remove: national government >> umbrella sciences institute that I forget the name of >> Loongson Corp.) so they have practically unlimited resources if the Chinese government decided to prioritize Loongson as the internal national computer architecture. Hire more people for the corporation directly, push a propaganda campaign through to get patriotic hackers to work on a Loongson centric software infrastructure as a duty. If they did that, they could probably accelerate development architecture performance and support considerably. It would still take years simply because of how long it takes to bring a finished product to the market plus create an acceptable end user software experience, just not as many. I doubt that'll actually happen, but it is at least plausible should international relations deteriorate enough.

I used to find MIPS an interesting architecture to play around with. Most of the "big/real" (not my term, I never considered x86 as "toys" even back then) systems I played with back in the early 90s were MIPS (early Ultrix DECStations & SGI) based Unix machines. I was employed to write compute programs for some physicists (in Fortran of course). Would be kinda neat to have access to a relatively modern MIPS evolution like Loongson to pick around at, but they aren't something you can pick up for cheap at Best Buy in the US. Emulation just isn't as fun to me. I like to poke at real hardware, with all its warts.

[szymon_g]

is there any hardware based on it available to easily(-ish) buy off ebay etc?

[xen0n]

is there any hardware based on it available to easily(-ish) buy off ebay etc?

I've heard several people in EU/NA have got their LoongArch boards via AliExpress, but I'd suggest waiting until the next generation that's Loongson 3A6000. The 3A5000 (and its server-class equivalent 3C5000/3C5000L) is basically a 3A4000 with modded decoder stage so it recognizes LoongArch instead of (also modded) MIPS64r5, and that micro-architecture is dated and not very competitive performance-wise, a bit like my 2011 laptop with quad-core Core i7. The 3A6000 board instead feels very much like some quad-core x86 from 2018 or so, so I'd really recommend waiting until such models are released...

[enihcam]

This is all true. However, Loongson Corp. is pretty much funded by the PRC state (through at one remove: national government >> umbrella sciences institute that I forget the name of >> Loongson Corp.)

just like intel is still being funded by the state, right?

[NovenTheHero]

just like intel is still being funded by the state, right?

I say that this is very much a false equivalence.