On Thu, Dec 10, 2009 at 11:02 PM, Stefan Strasser <strasser@uni-bremen.de> wrote:
what I find surprising is the performance of MySQL for small transactions. while my library spends 90% of its time waiting for the disk to flush a transaction, InnoDB uses 100% CPU and barely uses any disk.
but, including all the SQL parsing, running 10000 small transactions takes about 4 seconds on my system. while this little code snippet alone takes 6-7 seconds, with much more disk usage:
int fd=open("test.bin",O_RDWR | O_CREAT | O_NOATIME); for(int c=0;c<10000;++c){ write(fd,&data,1); fsync(fd); }
neither fdatasync(), O_WRONLY, nor writing whole sectors instead of just a byte make a difference. MySQL also claims to use fsync on the page you linked.
do you have an explanation for this?
You should be comparing using msync (not fsync). Also, the high cpu usage might be caused by innodb gzip compression regards