Re: [opensuse-kernel] kernel-preempt in Kernel:stable WAS: Re: Jump: merge Leap and SLE kernel configs
Am Dienstag, 15. September 2020, 16:23:06 CEST schrieb Michal Hocko:
On Fri 04-09-20 14:25:57, Michal Kubecek wrote:
On Fri, Sep 04, 2020 at 01:35:43PM +0200, Hans-Peter Jansen wrote:
Am Donnerstag, 3. September 2020, 19:14:56 CEST schrieb Takashi Iwai:
e.g. about the preemption, we can follow the SLE pattern, namely,
kernel-default = PREEMPT_NONE (typically server usage) and kernel-preempt = PREEMPT_FULL (typically desktop usage)
instead of a single kernel flavor for all.
Does that mean, we can expect another kernel flavor (-preempt) in Kernel:stable soon? This would be *much* appreciated, of course.
Not directly as this is about Leap where we already have preempt flavor (even if only for x86_64 and aarch64) and the only difference is that kernel-default has PREEMPT_VOLUNTARY on Leap and PREEMPT_NONE on SLE. So the only question is if we switch Leap kernel-default to PREEMPT_NONE in openSUSE-15.3 to match SLE15-SP3.
I would go with PREEMPT_NONE for default and PREEMPT_FULL for -preempt. This would be to cover the major usecases for throughput and latency.
On the other hand, having two flavors also in Tumbleweed would IMHO make sense but I would suggest to wait until we decide what settings we use in 15.3 and do the same in Tumbleweed then.
Yes having two kernel flavors makes sense.
Just in case, anybody wants to try this with Tumbleweed, here's a build for you: https://build.opensuse.org/project/monitor/home:frispete:kernel Target: openSUSE_Tumbleweed_Kernel_stable This build adheres to the official workflow. As such it is based on a fork of https://github.com/openSUSE/kernel-source and just contains two minimally invasive changes: https://github.com/frispete/kernel-source/tree/stable. One adds another flavor: {arm,x86_}64-preempt and raises HZ to 1000 for x86_64. The other switches to PREEMPT_NONE and NO_HZ_FULL for {arm,x86_}64-default. The preempt flavor behaves fine on a couple of very different (desktop/ notebook) systems so far. I expected a somewhat higher battery drain on my (customized) Lenovo X1C Gen1, but couldn't perceive such (while not measured). More the contrary, the 5.8 series appears to be improved in this area. Nice. I would like to discuss the NO_HZ_FULL versus NO_HZ_IDLE impacts for -default (the designated server choice). Anybody remember the reasons for this choice on SLE15-SP3-default? If you try these builds, please send some feedback on the results. Cheers, Pete -- To unsubscribe, e-mail: opensuse-kernel+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-kernel+owner@opensuse.org
On Wed, 16 Sep 2020 12:12:06 +0200, Hans-Peter Jansen wrote:
Am Dienstag, 15. September 2020, 16:23:06 CEST schrieb Michal Hocko:
On Fri 04-09-20 14:25:57, Michal Kubecek wrote:
On Fri, Sep 04, 2020 at 01:35:43PM +0200, Hans-Peter Jansen wrote:
Am Donnerstag, 3. September 2020, 19:14:56 CEST schrieb Takashi Iwai:
e.g. about the preemption, we can follow the SLE pattern, namely,
kernel-default = PREEMPT_NONE (typically server usage) and kernel-preempt = PREEMPT_FULL (typically desktop usage)
instead of a single kernel flavor for all.
Does that mean, we can expect another kernel flavor (-preempt) in Kernel:stable soon? This would be *much* appreciated, of course.
Not directly as this is about Leap where we already have preempt flavor (even if only for x86_64 and aarch64) and the only difference is that kernel-default has PREEMPT_VOLUNTARY on Leap and PREEMPT_NONE on SLE. So the only question is if we switch Leap kernel-default to PREEMPT_NONE in openSUSE-15.3 to match SLE15-SP3.
I would go with PREEMPT_NONE for default and PREEMPT_FULL for -preempt. This would be to cover the major usecases for throughput and latency.
On the other hand, having two flavors also in Tumbleweed would IMHO make sense but I would suggest to wait until we decide what settings we use in 15.3 and do the same in Tumbleweed then.
Yes having two kernel flavors makes sense.
Just in case, anybody wants to try this with Tumbleweed, here's a build for you: https://build.opensuse.org/project/monitor/home:frispete:kernel
Target: openSUSE_Tumbleweed_Kernel_stable
This build adheres to the official workflow. As such it is based on a fork of https://github.com/openSUSE/kernel-source and just contains two minimally invasive changes: https://github.com/frispete/kernel-source/tree/stable.
One adds another flavor: {arm,x86_}64-preempt and raises HZ to 1000 for x86_64.
The other switches to PREEMPT_NONE and NO_HZ_FULL for {arm,x86_}64-default.
The preempt flavor behaves fine on a couple of very different (desktop/ notebook) systems so far. I expected a somewhat higher battery drain on my (customized) Lenovo X1C Gen1, but couldn't perceive such (while not measured). More the contrary, the 5.8 series appears to be improved in this area. Nice.
I would like to discuss the NO_HZ_FULL versus NO_HZ_IDLE impacts for -default (the designated server choice). Anybody remember the reasons for this choice on SLE15-SP3-default?
NO_HZ_FULL has been the choice for SLE kernel since many years ago. IIRC, it's already on 4.4-based kernels. Looking at the git log, it was marked as FATE#318845. A good news is that it implies, at least, that NO_HZ_FULL can't be too bad for servers; otherwise SUSE must have received lots of complaints. But, it's better to measure the actual differences somehow in a real test environment, of course. Takashi -- To unsubscribe, e-mail: opensuse-kernel+unsubscribe@opensuse.org To contact the owner, e-mail: opensuse-kernel+owner@opensuse.org
participants (2)
-
Hans-Peter Jansen -
Takashi Iwai