Intel Skylake Graphics: Windows 10 vs. Ubuntu Linux Performance

[jf33]

Still no one can beat mesa drivers when it comes to correctness.

[gamerk2]

So in a Microsoft environment where few applications use openGL it still runs circles around Linux on this Intel GPU?

Is it just the weakness of the Linux driver or something to do with the OS?

I'm convinced the way the Linux scheduler manages threads is partially to blame. When running heavy benchmarks, you want the scheduler biased to running that application at all times, even if background tasks get starved. Windows does this, Linux does not. I'm convinced the GPU/Application thread(s) are getting swapped out, which could kill performance by well over 50%.

Unfortunately, Linux is short on low-level debug tools that can actually give a user that degree of visibility; Linux seriously needs something akin to GPUView, if it really wants to get serious about gaming results.

[Eero]

Michael, if you want results that are comparable between Windows and Linux, you should run fullscreen tests only in monitor's native resolution. Or verify the resolution actually used by the test with "xwininfo" (run e.g. from ssh console).

Many tests don't behave similarly on Windows and Linux (and neither does the windowing system) if requested resolution differs from the monitor's native one. I.e. use FullHD monitor, if you want to run tests in fullscreen FullHD resolution, and if you want to compare them in some other resolution, run them in windowed mode.

Besides Xonotic, GpuTest results seem also bogus. Several of the Linux results are about 1/4 of what they should be. E.g. when Volplosion is run correctly on Skylake, it's faster with Linux/Mesa than with Windows.

Mesa supports GpuTest GiMark just fine as that doesn't use any new GL features, issue is in how Phoronix test-suite runs it: GiMark is buggy, you need to specify the GL version for Mesa to accept its output. Any GL 3.x version is OK, e.g. "MESA_GL_VERSION_OVERRIDE=3.1". I've also seen Windows to misrender that test, but I don't know whether that affects performance and whether that's a problem in the test or Windows driver.

Note: Above isn't only issue with GpuTest, e.g. test duration directly multiplies the resulting test score...

[johnc]

I'm convinced the way the Linux scheduler manages threads is partially to blame. When running heavy benchmarks, you want the scheduler biased to running that application at all times, even if background tasks get starved. Windows does this, Linux does not. I'm convinced the GPU/Application thread(s) are getting swapped out, which could kill performance by well over 50%.

We already have benchmarks of games that are CPU bottlenecked where Linux is ahead, when using drivers from non-idiotic companies.

[Eero]

GpuTest tessalation test can be run with Mesa branch from here: http://cgit.freedesktop.org/~kwg/mesa/log/?h=tessquash

By prefixing the test with MESA_GL_VERSION_OVERRIDE=4.0.