CentOS 7 | microcode | dramatic slowdown

General support questions
Post Reply
egc
Posts: 4
Joined: 2019/09/26 17:06:51

CentOS 7 | microcode | dramatic slowdown

Post by egc » 2019/10/07 13:33:19

So, running CentOS 7 on a fairly high-end box -- applied the latest microcode patch, and the system is palpably slower. I did a little digging, and based on this thread

viewtopic.php?t=65767

if I run cat /sys/kernel/debug/x86/*enabled , I should see current status of various files (like ibpb, ibrs, pti_enabled)., in theory, three, but when I run the command I get the following:

Code: Select all

> cat /sys/kernel/debug/x86/*enabled
1
0
1
1
3
So not only don't I get three (I get five), but they're not all binary -- the last one is '3'.

Suggestions? Pointers to the obvious?

Thanks in advance...
Last edited by egc on 2019/10/07 21:24:48, edited 2 times in total.

chemal
Posts: 776
Joined: 2013/12/08 19:44:49

Re: entOS 7 | microcode | dramatic slowdown

Post by chemal » 2019/10/07 20:53:11

Read /usr/share/doc/kernel-doc-3.10.0/Documentation/spec_ctrl.txt from the current kernel-doc package instead of old threads.

What does "palpably slower" exactly mean?

egc
Posts: 4
Joined: 2019/09/26 17:06:51

Re: entOS 7 | microcode | dramatic slowdown

Post by egc » 2019/10/07 21:15:32

chemal wrote:
2019/10/07 20:53:11
Read /usr/share/doc/kernel-doc-3.10.0/Documentation/spec_ctrl.txt from the current kernel-doc package instead of old threads.
Will do -- thanks!
What does "palpably slower" exactly mean?
10-15% slower on certain benchmarks - all numerically-based. Perhaps 10-15% is 'acceptable', but propogating that to jobs that takes hours to finish is quite a bit.

egc
Posts: 4
Joined: 2019/09/26 17:06:51

Re: entOS 7 | microcode | dramatic slowdown

Post by egc » 2019/10/07 21:20:40

chemal wrote:
2019/10/07 20:53:11
Read /usr/share/doc/kernel-doc-3.10.0/Documentation/spec_ctrl.txt
Didn't realize you have to install kernel-doc (not installed by default, at least on my install).

chemal
Posts: 776
Joined: 2013/12/08 19:44:49

Re: CentOS 7 | microcode | dramatic slowdown

Post by chemal » 2019/10/07 21:30:53

Here is what RH says about performance impacts: https://access.redhat.com/articles/3307751

10-15% seems excessive for "numerically-based" benchmarks. What benchmarks exactly?

Edit: 7.7 defaults to retpoline on Skylake (7.6 didn't), but only for new installs. If you updated 7.6 to 7.7 you have to put "spectre_v2=retpoline" on the kernel command line if you want the new default.

You can also experiment with "mitigations=off". If this doesn't help you have some other problem.

egc
Posts: 4
Joined: 2019/09/26 17:06:51

Re: CentOS 7 | microcode | dramatic slowdown

Post by egc » 2019/10/07 22:17:15

chemal wrote:
2019/10/07 21:30:53
Here is what RH says about performance impacts: https://access.redhat.com/articles/3307751

10-15% seems excessive for "numerically-based" benchmarks. What benchmarks exactly?


My own number-crunching code for the work I do (lots of matrix stuff, openMP calls, etc). I'm going to run the polyhedron benchmarks next -- for all I know, its a compiler switch that doesn't play nice anymore...
Edit: 7.7 defaults to retpoline on Skylake (7.6 didn't), but only for new installs. If you updated 7.6 to 7.7 you have to put "spectre_v2=retpoline" on the kernel command line if you want the new default.

You can also experiment with "mitigations=off". If this doesn't help you have some other problem.
Easy enough to try the latter -- thanks!

Post Reply