Re: [PATCH v3 0/3] Introduce per-task latency_nice for scheduler hints

From: chris hyser
Date: Tue Feb 18 2020 - 18:01:05 EST

On 2/17/20 3:57 AM, Parth Shah wrote:

On 1/16/20 5:32 PM, Parth Shah wrote:
This is the 3rd revision of the patch set to introduce
latency_{nice/tolerance} as a per task attribute.

The previous version can be found at:

Changes in this revision are:
v2 -> v3:
- This series changes the longer attribute name to "latency_nice" as per
the comment from Dietmar Eggemann
v1 -> v2:
- Addressed comments from Qais Yousef
- As per suggestion from Dietmar, moved content from newly created
include/linux/sched/latency_tolerance.h to kernel/sched/sched.h
- Extend sched_setattr() to support latency_tolerance in tools headers UAPI

This patch series introduces a new per-task attribute latency_nice to
provide the scheduler hints about the latency requirements of the task [1].

Latency_nice is a ranged attribute of a task with the value ranging
from [-20, 19] both inclusive which makes it align with the task nice

The value should provide scheduler hints about the relative latency
requirements of tasks, meaning the task with "latency_nice = -20"
should have lower latency requirements than compared to those tasks with
higher values. Similarly a task with "latency_nice = 19" can have higher
latency and hence such tasks may not care much about latency.

The default value is set to 0. The usecases discussed below can use this
range of [-20, 19] for latency_nice for the specific purpose. This
patch does not implement any use cases for such attribute so that any
change in naming or range does not affect much to the other (future)
patches using this. The actual use of latency_nice during task wakeup
and load-balancing is yet to be coded for each of those usecases.

As per my view, this defined attribute can be used in following ways for a
some of the usecases:
1 Reduce search scan time for select_idle_cpu():
- Reduce search scans for finding idle CPU for a waking task with lower
latency_nice values.

2 TurboSched:
- Classify the tasks with higher latency_nice values as a small
background task given that its historic utilization is very low, for
which the scheduler can search for more number of cores to do task
packing. A task with a latency_nice >= some_threshold (e.g, == 19)
and util <= 12.5% can be background tasks.

3 Optimize AVX512 based workload:
- Bias scheduler to not put a task having (latency_nice == -20) on a
core occupying AVX512 based workload.

Series Organization:
- Patch 1: Add new attribute latency_nice to task_struct.
- Patch 2: Clone parent task's attribute to the child task on fork
- Patch 3: Add support for sched_{set,get}attr syscall to modify
latency_nice of the task

The patch series can be applied on tip/sched/core at the
commit 804d402fb6f6 ("sched/rt: Make RT capacity-aware")

[1]. Usecases for the per-task latency-nice attribute,
[2]. Task Latency-nice, "Subhra Mazumdar",
[3]. Introduce per-task latency_tolerance for scheduler hints,

Parth Shah (3):
sched: Introduce latency-nice as a per-task attribute
sched/core: Propagate parent task's latency requirements to the child
sched: Allow sched_{get,set}attr to change latency_nice of the task

include/linux/sched.h | 1 +
include/uapi/linux/sched.h | 4 +++-
include/uapi/linux/sched/types.h | 19 +++++++++++++++++++
kernel/sched/core.c | 21 +++++++++++++++++++++
kernel/sched/sched.h | 18 ++++++++++++++++++
tools/include/uapi/linux/sched.h | 4 +++-
6 files changed, 65 insertions(+), 2 deletions(-)

Its been a long time and few revisions since the beginning of the
discussion around the latency-nice. Hence thought of asking if there is/are
any further work that needs to be done for adding latency-nice attribute or
am I missing any piece in here?

All, I was asked to take a look at the original latency_nice patchset. First, to clarify objectives, Oracle is not interested in trading throughput for latency. What we found is that the DB has specific tasks which do very little but need to do this as absolutely quickly as possible, ie extreme latency sensitivity. Second, the key to latency reduction in the task wakeup path seems to be limiting variations of "idle cpu" search. The latter particularly interests me as an example of "platform size based latency" which I believe to be important given all the varying size VMs and containers.

Parth, I've been using your v3 patchset as the basis of an investigation into the measurable effects of short-circuiting this search. I'm not quite ready to put anything out, but the patchset is working well. The only feedback I have is that currently non-root can set the value negative which is inconsistent with 'nice' and I would think a security hole.