I wrote: > 3) This topic has been discussed in the glibc bug > https://sourceware.org/bugzilla/show_bug.cgi?id=13701 > where I have raised my voice for a writer-preferring implementation.
The rwlock implementation being not writer-preferring appears to be only one contributing factor. I did "time ./test-pthread-rwlock" * in virtual machines with 8 CPUs: - Ubuntu 22.04 60 sec (4 CPUs: 15 sec) - musl libc on Ubuntu 22.04 0.3 sec - Alpine Linux 0.3 sec - FreeBSD 14 1.4 sec - OpenBSD 7.5 2.5 sec - Solaris 11.4 22 sec (4 CPUs: 20 sec) - Cygwin 33..35 sec (4 CPUs: 25..27 sec) - mingw 12 sec - MSVC 16 sec * outside VirtualBox: - Ubuntu 22.04 (12 CPUs) 6 sec - macOS 12 3.6 sec - AIX 7.1 4 sec Intermediate results: * While gl_cv_pthread_rwlock_rdlock_prefer_writer is - yes on macOS, FreeBSD, NetBSD, OpenBSD <= 6.0, Solaris, Cygwin, - no on glibc, musl libc, OpenBSD >= 6.5, AIX, Android, only glibc has the problem; musl libc, OpenBSD 7.5, AIX don't. * The first two measurements were done on the same Ubuntu 22.04 VM. This hints that kernel customization (scheduler, policies etc.) are not the cause — otherwise musl libc on Ubuntu 22.04 would have the same problem. * I don't understand why the same Ubuntu 22.04 outside VirtualBox works fine (6 seconds), whereas inside VirtualBox it's 60 seconds. But it's not entirely a VirtualBox or virtualization problem, since in Paul's case the system is running on the naked harware (no virtualization). So: - It seems to be a glibc problem. - But a large amount of CPU time is spent in the kernel. Which performance tools exist that can help understand the situation across the glibc / kernel boundary? Bruno
