There is a long history of distro people, system administrators, and
application owners tuning the CFS settings in /proc/sys, which are now
in debugfs. Looking at what these settings actually did, it ended up
boiling down to changing the likelihood of task preemption, or
disabling it by setting the wakeup_granularity_ns to more than half of
the latency_ns. The other settings didn't really do much for
performance.
In other words, some our workloads benefit by having long running tasks
preempted by tasks handling short running requests, and some workloads
that run only short term requests which benefit from never being preempted.
This leads to a few observations and ideas:
- Different workloads want different policies. Being able to configure
the policy per workload could be useful.
- A workload that benefits from not being preempted itself could still
benefit from preempting (low priority) background system tasks.
- It would be useful to quickly (and safely) experiment with different
policies in production, without having to shut down applications or reboot
systems, to determine what the policies for different workloads should be.
- Only a few workloads are large and sensitive enough to merit their own
policy tweaks. CFS by itself should be good enough for everything else,
and we probably do not want policy tweaks to be a replacement for anything
CFS does.
This leads to BPF hooks, which have been successfully used in various
kernel subsystems to provide a way for external code to (safely)
change a few kernel decisions. BPF tooling makes this pretty easy to do,
and the people deploying BPF scripts are already quite used to updating them
for new kernel versions.
This patchset aims to start a discussion about potential applications of BPF
to the scheduler. It also aims to land some very basic BPF infrastructure
necessary to add new BPF hooks to the scheduler, a minimal set of useful
helpers, corresponding libbpf changes, etc.
Our very first experiments with using BPF in CFS look very promising. We're
at a very early stage, however already have seen a nice latency and ~1% RPS
wins for our (Facebook's) main web workload.
As I know, Google is working on a more radical approach [2]: they aim to move
the scheduling code into userspace. It seems that their core motivation is
somewhat similar: to make the scheduler changes easier to develop, validate
and deploy. Even though their approach is different, they also use BPF for
speeding up some hot paths. I think the suggested infrastructure can serve
their purpose too.
An example of an userspace part, which loads some simple hooks is available
here [3]. It's very simple, provided only to simplify playing with the provided
kernel patches.
[1] c722f35b513f ("sched/fair: Bring back select_idle_smt(), but differently")
[2] Google's ghOSt: https://linuxplumbersconf.org/event/11/contributions/954/
[3] https://github.com/rgushchin/atc
Roman Gushchin (6):
bpf: sched: basic infrastructure for scheduler bpf
bpf: sched: add convenient helpers to identify sched entities
bpf: sched: introduce bpf_sched_enable()
sched: cfs: add bpf hooks to control wakeup and tick preemption
libbpf: add support for scheduler bpf programs
bpftool: recognize scheduler programs
include/linux/bpf_sched.h | 53 ++++++++++++
include/linux/bpf_types.h | 3 +
include/linux/sched_hook_defs.h | 4 +
include/uapi/linux/bpf.h | 25 ++++++
kernel/bpf/btf.c | 1 +
kernel/bpf/syscall.c | 21 ++++-
kernel/bpf/trampoline.c | 1 +
kernel/bpf/verifier.c | 9 ++-
kernel/sched/Makefile | 1 +
kernel/sched/bpf_sched.c | 138 ++++++++++++++++++++++++++++++++
kernel/sched/fair.c | 27 +++++++
scripts/bpf_doc.py | 2 +
tools/bpf/bpftool/common.c | 1 +
tools/bpf/bpftool/prog.c | 1 +
tools/include/uapi/linux/bpf.h | 25 ++++++
tools/lib/bpf/libbpf.c | 27 ++++++-
tools/lib/bpf/libbpf.h | 4 +
tools/lib/bpf/libbpf.map | 3 +
18 files changed, 341 insertions(+), 5 deletions(-)
create mode 100644 include/linux/bpf_sched.h
create mode 100644 include/linux/sched_hook_defs.h
create mode 100644 kernel/sched/bpf_sched.c
--
2.31.1
