Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]
Groups > linux.debian.kernel > #80695
| From | Emanuele Rocca <ema@debian.org> |
|---|---|
| Newsgroups | linux.debian.kernel |
| Subject | Re: CONFIG_PREEMPT_DYNAMIC=y? |
| Date | 2023-10-20 20:40 +0200 |
| Message-ID | <Hr2zD-heZn-9@gated-at.bofh.it> (permalink) |
| References | <HmEPf-eqdx-7@gated-at.bofh.it> |
| Organization | linux.* mail to news gateway |
Hi, On 2023-10-08 06:20, Johannes Schauer Marin Rodrigues wrote: > I'm wondering whether it'd make sense for the Debian kernel to be built with > CONFIG_PREEMPT_DYNAMIC=y. On amd64 we already have PREEMPT_DYNAMIC=y, so I thought of running some benchmarks on my Ryzen system to evaluate the performance overhead it introduces. To my surprise, I found that in Debian you get a 20% (!) performance improvement on kernel-heavy workloads simply by setting PREEMPT_DYNAMIC to 'n'. I brought up the topic on LKML [1], and it turns out that what is really eating performance for us is DEBUG_PREEMPT=y. There is a performance penalty in PREEMPT_DYNAMIC=y as well, but it's much smaller, around 3 or 4%. The Kconfig help entry for DEBUG_PREEMPT also clearly mentions the performance overhead: This option has potential to introduce high runtime overhead, depending on workload as it triggers debugging routines for each this_cpu operation. It should only be used for debugging purposes. Regardless of what we are going to decide about PREEMPT_DYNAMIC, we surely have to disable DEBUG_PREEMPT. I opened [2] for that. [1] https://lore.kernel.org/lkml/ZTJFA_Ac6nWawIHb@ariel/ [2] https://salsa.debian.org/kernel-team/linux/-/merge_requests/874
Back to linux.debian.kernel | Previous | Next — Previous in thread | Find similar | Unroll thread
CONFIG_PREEMPT_DYNAMIC=y? Johannes Schauer Marin Rodrigues <josch@debian.org> - 2023-10-08 18:30 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Diederik de Haas <didi.debian@cknow.org> - 2023-10-08 19:30 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Emanuele Rocca <ema@debian.org> - 2023-10-10 12:10 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Johannes Schauer Marin Rodrigues <josch@debian.org> - 2023-10-10 12:20 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Emanuele Rocca <ema@debian.org> - 2023-10-10 12:40 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Marc Haber <mh+debian-kernel@zugschlus.de> - 2023-10-10 15:50 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Diederik de Haas <didi.debian@cknow.org> - 2023-10-10 14:00 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Vincent Blut <vincent.debian@free.fr> - 2023-10-10 14:20 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Diederik de Haas <didi.debian@cknow.org> - 2023-10-10 15:20 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Emanuele Rocca <ema@debian.org> - 2023-10-10 15:00 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Diederik de Haas <didi.debian@cknow.org> - 2023-10-10 23:30 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Diederik de Haas <didi.debian@cknow.org> - 2023-10-17 20:00 +0200
Re: CONFIG_PREEMPT_DYNAMIC=y? Emanuele Rocca <ema@debian.org> - 2023-10-20 20:40 +0200
csiph-web