[Bug] [KNL] ~5x slowdown of application when attaching by perf on KNL

Bug #1730782 reported by quanxian on 2017-11-08
6
This bug affects 1 person
Affects Status Importance Assigned to Milestone
intel
Undecided
Unassigned
linux (Ubuntu)
Medium
Canonical Kernel Team
Bionic
Medium
Canonical Kernel Team

Bug Description

Description:

Workload slows down for default sampling cycles in KNL. There are some processing in perf that's single threaded and bottlenecked when number of cores really go up.
A second issue is that things slow down when we issue
perf record – workload
vs
perf record -a – workload

Patch is submitted for review

https://www.spinics.net/lists/kernel/msg2552460.html
https://www.spinics.net/lists/kernel/msg2552461.html
https://www.spinics.net/lists/kernel/msg2552462.html
https://www.spinics.net/lists/kernel/msg2552463.html

Target kernel: 4.17
Target Release: 18.10

tags: added: kernel
information type: Proprietary → Public
Changed in linux (Ubuntu Bionic):
status: New → Triaged
importance: Undecided → Medium
assignee: nobody → Canonical Kernel Team (canonical-kernel-team)
tags: added: bionic kernel-da-key

A v9 was posted on August 31st, but there is still no progress upstream. Meanwhile, Peter Zijlstra merged 0d3d73aac2ff05c78387aa9dcc2c8aa3804405e7 ("perf/core: Rewrite event timekeeping"), which looks like would have helped with this problem.

Can you confirm if the latest bionic kernel, which has this commit, scales better? If not, what is the plan of pushing this forward upstream?

Cascardo.

quanxian (quanxian-wang) on 2018-05-01
description: updated
tags: added: intel-kernel-18.10
Changed in intel:
status: New → Won't Fix
quanxian (quanxian-wang) wrote :

the patch of this bug is under review in upstream. Once update is official, we will provide patch commit id in this bug. This bug is for tracking. Thanks for your support.

quanxian (quanxian-wang) on 2018-10-24
Changed in intel:
status: Won't Fix → New
tags: added: intel-kernel-19.04
removed: intel-kernel-18.10
description: updated
quanxian (quanxian-wang) on 2019-02-28
description: updated
quanxian (quanxian-wang) wrote :

has been fixed in 4.17

Changed in intel:
status: New → Fix Released

Hi @quanxian. What commits are these? Do you want these on bionic 4.15?

Cascardo.

quanxian (quanxian-wang) on 2019-03-04
description: updated
description: updated
quanxian (quanxian-wang) wrote :

No special request for bionic. Just for 18.10 will be enough.

only one is provided. Other 3 patches are not found. Upstream tested it that has been fixed.

https://www.spinics.net/lists/kernel/msg2552460.html

commit 8e1a2031e4b556b01ca53cd1fb2d83d811a6605b
Author: Alexey Budankov <email address hidden>
Date: Fri Sep 8 11:47:03 2017 +0300

perf/cor: Use RB trees for pinned/flexible groups

Change event groups into RB trees sorted by CPU and then by a 64bit
index, so that multiplexing hrtimer interrupt handler would be able
skipping to the current CPU's list and ignore groups allocated for the
other CPUs.

quanxian (quanxian-wang) wrote :

no more work to be done. 18.10 has contained the patch.

git describe 8e1a2031e4b556b01ca53cd1fb2d83d811a6605b
v4.16-rc4-159-g8e1a2031e4b5

To post a comment you must log in.
This report contains Public information  Edit
Everyone can see this information.

Other bug subscribers