Re: [boost] [threads] read_write_mutex performance

From: Caleb Epstein (caleb.epstein_at_[hidden])
Date: 2005-04-25 10:34:10

On 4/24/05, Peter Dimov <pdimov_at_[hidden]> wrote:

> What are the design goals of read_write_mutex? (Lock proliferation aside for
> a moment.)

Attached is a copy of the test program modified to use the ACE library
and to use n_writers = 1.  Performance of the two cases is much
closer, though the read-write case still suffers compared to the
simple mutex case and is really platform-dependant.  Here are my
results (where rwlock-ace is the simple case and rwlock-ace-rw is the
read/write case):

Linux/g++:
./rwlock-ace  3.90s user 14.54s system 326% cpu 5.648 total
./rwlock-ace-rw  6.94s user 26.57s system 226% cpu 14.776 total

Solaris/Sun C++:
./rwlock-ace-sun  7.64s user 50.66s system 691% cpu 8.427 total
./rwlock-ace-sun-rw  34.19s user 250.70s system 465% cpu 1:01.18 total

I'd say the design of the Boost.Threads read/write mutex is
sub-optimal compared to the one on ACE.

-- 
Caleb Epstein
caleb dot epstein at gmail dot com