Comments on: Comparison: Lockless programming with atomics in C++ 11 vs. mutex and RW-locks

By: Willi Goesgens

Willi Goesgens — Wed, 11 Mar 2020 09:14:33 +0000

Afair it was seconds.
the gist with the source probably shows: https://gist.github.com/13abylon/523685ef6af5bde24dc3

By: William Baxter

William Baxter — Tue, 10 Mar 2020 22:09:42 +0000

What units are the table in? Microseconds?

By: jsteemann

jsteemann — Mon, 14 Dec 2015 08:57:00 +0000

In reply to Marcus_FreeAB. Thanks, this looks indeed promising!

By: Marcus_FreeAB

Marcus_FreeAB — Mon, 14 Dec 2015 08:33:00 +0000

I ran the tests on an Intel Xeon E5-2630 running Windows 7 and got the following results:

VS 2012:

Duration: 0.012001
Duration: 205.906
Duration: 1.25512
Duration: 1.46515
Duration: 1.63016
Duration: 1.56016
Duration: 1.79218
Duration: 1.78818
Duration: 1.77418
Duration: 1.73417
Duration: 1.64416
Duration: 1.66317
Duration: 1.78618

VS 2015

Duration: 0.00999999
Duration: 2.52
Duration: 1
Duration: 1.17
Duration: 0.03
Duration: 0.00999999
Duration: 0.00999999
Duration: 0.02
Duration: 0.03
Duration: 0.00999999
Duration: 0.00999999
Duration: 0.02
Duration: 0.02

Note that the std::mutex is “fixed” and the speed overall is better in 2015.

Conclusion: Upgrade to VS2015 NOW!

By: Wilfried Gösgens

Wilfried Gösgens — Thu, 19 Feb 2015 14:54:00 +0000

In reply to L.A.J.W.. try rebooting the machine with a live linux like knoppix or GRML.

By: L.A.J.W.

L.A.J.W. — Tue, 17 Feb 2015 23:10:00 +0000

In reply to Wilfried Gösgens.

Testing different platforms is important, so we know what’s best overall.
Tested on Windows 8.1 x64 on “antique” overclocked i7 860. VS 2013 Release Build x86 (4 core, 8 threads, 3.6 Ghz flat):
#0 0.03 // Depressing
#1 1.43 // Intriguing
#2 0.20 // More intriguing
#3 0.29
#4 0.19 // Wow
#5 0.19
Now we only need to test AMD on Linux. I don’t have one on my hands right now unfortunately.

By: Yiannis Papadopoulos

Yiannis Papadopoulos — Tue, 17 Feb 2015 18:24:00 +0000

Just as I side-note, most compilers at don’t really optimize for std::memory_order_consume and simply assume it’s the same as std::memory_order_acquire. See also http://preshing.com/20140709/the-purpose-of-memory_order_consume-in-cpp11/

By: Wilfried Gösgens

Wilfried Gösgens — Tue, 17 Feb 2015 11:24:00 +0000

In reply to L.A.J.W..

The main intend of this comparison was to prove the time investment into
embracing atomics worthwhile then getting the prove how bad windows
performs on which cpu architecture.
I didn’t have another windows box available at that time. Since the source is available, I would love to see your testresults from other machines. Grml.org offers nice slink live cds to do the linux comparison.

By: Iron Bug

Iron Bug — Tue, 17 Feb 2015 09:54:00 +0000

In reply to Wilfried Gösgens. Yes, on Windows it was always a problem. I used boost conditional variables (as quite fast method among other cross-platform solutions) and asm-based spinlocks for synchronizing threads back in times when I designed software for Windows. Now I deal mainly with Linux or Unix but still remember all the problems with timers and synchronizing in Windows and cross-platform decisions. And we also use ACE library: it's cross-platform, offers locks, RW-locks that often work faster than boost, although having poor documentation.

By: Wilfried Gösgens

Wilfried Gösgens — Tue, 17 Feb 2015 09:46:00 +0000

In reply to Iron Bug.

I found the locklessinc.com too after doing these tests; They offer a nice pthread.h for windows which may also work nice as porting layer:
http://locklessinc.com/articles/pthreads_on_windows/