Bug 1976267

Summary: memset performance differences between Ivy Bridge and Cascade Lake CPU families
Product: Red Hat Enterprise Linux 8 Reporter: Sebastien Aime <saime>
Component: glibcAssignee: DJ Delorie <dj>
Status: CLOSED INSUFFICIENT_DATA QA Contact: qe-baseos-tools-bugs
Severity: medium Docs Contact:
Priority: medium    
Version: 8.2CC: ashankar, codonell, dj, fweimer, mnewsome, pfrankli, sipoyare
Target Milestone: betaKeywords: Bugfix, Triaged
Target Release: ---   
Hardware: x86_64   
OS: Linux   
Whiteboard:
Fixed In Version: Doc Type: If docs needed, set a value
Doc Text:
Story Points: ---
Clone Of: Environment:
Last Closed: 2021-10-18 07:43:52 UTC Type: Bug
Regression: --- Mount Type: ---
Documentation: --- CRM:
Verified Versions: Category: ---
oVirt Team: --- RHEL 7.3 requirements from Atomic Host:
Cloudforms Team: --- Target Upstream Version:
Embargoed:
Attachments:
Description Flags
memset systemtap probe script none

Description Sebastien Aime 2021-06-25 16:11:04 UTC
Description of problem:

An application performance benchmark has highlighted some differences in memset calls when the same code is run on Ivy Bridge CPU's or Cascade Lake CPU's.
The application is compiled on RHEL-7 on an Ivy Bridge system and then run on RHEL-8 systems equipped with Ivy Bridge and Cascade Lake CPU's.

Here is an example code that will highlight this performance difference:

#include <stdlib.h>

int main(int argc, char * argv[]) {
  const int s = atoi(argv[1]) + 0;
  int size = s * 1024 * 1024;
  char * p;
  for(int i = 0; i < s; ++i) {
    p = calloc(s, sizeof(int));
  }
  return 0;
}

This code is compiled on RHEL-7 with the Developer Toolset 8:

source /opt/rh/devtoolset-8/enable
gcc -g test_calloc.c -o test_calloc
valgrind --tool=callgrind --instr-atstart=yes test_calloc 10000

With valgrind one can observe the following:


- Ivy Bridge (Intel(R) Xeon(R) CPU E5-2690 v2) (glibc-2.28)
Total cycle estimation cost: 46 642 220
__memset_sse2_unaligned: 43 799 668

- Cascade Lake (Intel(R) Xeon(R) Gold 6246R) (glibc-2.28)
Total cycle estimation cost: 401 717 850
__memset_avx2_unaligned_erms: 398 874 712

--- edit ---

fixed incorrect values in the valgrind outputs.

Comment 10 DJ Delorie 2021-06-30 21:37:42 UTC
I built the test program on an E5-2640 and a 6258R (I couldn't find
exact matches, YMMV).  As a sanity check, I did this on both:

$ time ./test_calloc 100000

It took 3.5 seconds on the older cpu and 2.3 seconds on the new one,
which is in line with expectations.  Could the customer try this
sanity check?

You noted that the test case was built on RHEL 7.  In this particular
test case, this is mostly irrelevent since all the processing happens
within glibc (where calloc and memset live), and on RHEL 8 you'll be
using the RHEL 8 glibc.  I assume the customer knows this, but I'm
mentioning it for clarity.  RHEL 8 has many optimizations that RHEL 7
doesn't have, so a performance difference is expected.  Some of these
optimizations are CPU-specific, so differences between CPU families is
also expected.

If, however, you find that a RHEL 7 test runs faster than a RHEL 8
test on the *same* cpu, we consider that to be a regression.

RHEL 8 does include some ability to tune performance using tunables;
see
https://www.gnu.org/software/libc/manual/html_node/Hardware-Capability-Tunables.html
(note this is for the current upstream, not specifically 2.28; use
"info libc" for documentation specific to your installation, it's in
the Tunables chapter.  Important: it's "glibc.tune.*" for 2.28 and
"glibc.cpu.*" for upstream)

You want to play with the glibc.tune.hwcaps tunable, like this
(memset.stap attached):

$ dnf install systemtap-devel
$ stap -c './test_calloc 1000' memset.stap
$ export GLIBC_TUNABLES=glibc.tune.hwcaps=-AVX2_Usable
$ stap -c './test_calloc 1000' memset.stap

Again for clarity - please note also that running an application under
valgrind or one of its related tools (like callgrind) may affect
performance - valgrind is not a benchmark tool, it's a debugging tool.
Perf, likewise, is better for relative performance within an
application, than absolute performance across differing hardware,
since CPU clock speeds and insns/clock may vary.  Benchmarking
short-lived functions like memset is difficult, see "Practical
micro-benchmarking with 'ltrace' and 'sched'"[1].

[1] https://developers.redhat.com/blog/2016/03/11/practical-micro-benchmarking-with-ltrace-and-sched

Comment 13 DJ Delorie 2021-07-06 17:37:26 UTC
Created attachment 1798738 [details]
memset systemtap probe script

Comment 14 DJ Delorie 2021-07-06 17:43:09 UTC
Sorry!  It's attached now.