Xin LI wrote: > Hi, Marian, > > 在 2005-10-26三的 11:14 +0200,Marian Hettwer写道: > >>Hi Folks, >> >>I'm using RC1 on some of my test servers and workstations. So far, it >>runs very good (as in stable and fast) :) >>However, I thought of doing a small benchmark of MySQL in comparison to >>Linux 2.6.8 (Debian Sarge) and the results are kinda... shocking. > > [...] > >>Any Ideas where this huge difference comes from? Did I oversee some >>debugging features still active in RC1 ? > > > There are many factors that can affect the performance of MySQL. Would > you please post the output of "dumpfs /var | head -n 21" on your server? > (Replace /var with whatever mountpoint that your database is stored in). > > Some observations that I think can affect MySQL performance: > - Our malloc(3) implementation might be sub-optimal for the workload. > - Lack of some POSIX system call which was implemented in other ways > (e.g. we do not have fsyncdata yet) > - Some detailed implementation difference between our and Linux's > threading library, or MySQL itself. > > I think David (davidxu_at_) had worked on performance of MySQL and he may > have some suggestions. > make sure key_buffer_size is 256M or more, set it in /etc/my.cnf: [mysqld] port = 3306 socket = /tmp/mysql.sock skip-locking key_buffer_size = 256M also, after key buffer size is no longer a bottleneck factor, try to repalce microtime() with getmicrotime() in gettimeofday syscall: the syscall is in /sys/kern/kern_time.c, here is some benchmark results: http://people.freebsd.org/~davidxu/mysql/gettimeofday_use_getmicrotime.txt http://people.freebsd.org/~davidxu/mysql/gettimeofday_use_microtime.txt Here is the ktrace dump for mysql server 4.1.14: http://people.freebsd.org/~davidxu/mysql/mysql_ktrace.txt Check gettimeofday syscall, it follows every I/O syscall, I think our gettimeofday is tooooooo expensive, if we can directly get time from memory, the performance will be improved further. > BTW. What file system are you using on your Debian test? What's the > mount options? > > Cheers,Received on Thu Oct 27 2005 - 05:12:25 UTC
This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:38:46 UTC