Due to the fact that I have received many personal emails, I now publish my last benchmark. Isn't a real query bench, because I don't test massively multithreading queries, but I launch sequential queries to misurate the impact time of the tcp open/close connection. So into this test we can (maybe) misurate only the I/O tcp speed of mysql daemons.

The client access directly to libmysql.dll. The server version is the 3.23.51. I use a fantastic borland compiler.

I reduced the output of the select at around 40 bytes (select limit 1)... so we can misurate only the I/O subsystem throughtput...

Unfortunately I have a winxp pro copy on the pc where I have a linux 2.4 kernel, I think that a win2k server with kernel quantum table linear at 120 msec can do best results (~5-10% I suppose)... hovewer we can say that the mysql I/O under linux can be ~10-20% faster that the equivalent under xp version.

www.dellapasqua.com/Linux2.4_pooled_LAN.gif www.dellapasqua.com/Linux2.4_NOTpooled_LAN.gif www.dellapasqua.com/XPPRO_pooled.gif www.dellapasqua.com/XPPRO_NOTpooled.gif www.dellapasqua.com/localhost_pooled.gif www.dellapasqua.com/localhost_NOTpooled.gif

My thougths are: - a pooling manager it's obligatory under win2k-xp client-server if you use a localhost mysqlserver accessed directly by the libmysql.dll (to avoid the famous localhost nt bad behavior) - if you have a separate machine with mysql you can build a pooled connection manager only if you need massive performances (yahoo?) - if you use a middleware API to access mysql, more probably you already have a pooling object (as ADO for ms that it's autopooling)

Thats'all and take these numbers only as curiosity!

Greetings to all.

Roberto.

ps. if you want a more speedily mysql under NT it's necessary to change the code to use nt kernels primitives (iocomps, not blocking sockets...etc) ^_^