Pythian Group

I got the idea for this article when one of our clients complained that their server’s performance was degrading during business hours. They thought it was weird that at the same time, SQL Server would list more than 1200 connections on SQL server Activity Monitor.

The server hosts more than 50 databases that serve an ASP.NET application hosted on some servers in a web farm. These servers issue connections to the databases in a distributed manner to balance the web application load. I tried to discover what these connections were doing and to what databases they were connected. Connections grouped by database:

select  db_name(dbid) , count(*) 'connections count'
from master..sysprocesses
where spid > 50 and spid  @@spid
group by  db_name(dbid)
order by count(*) desc

This showed some databases having more than 300 connections associated with them.

What about logins used?

select  loginame , nt_username, count(*) 'Connections count'
from master..sysprocesses
where spid > 50 and spid  @@spid
group by  loginame , nt_username
order by count(*) desc

This showed a mix of windows domain accounts (those with values in column nt_username, e.g: domain\user) beside SQL authentication accounts (those with column nt_username empty, e.g: “sa”).

In order to reduce the number of times that new connections must be opened, applications may use connection pooling. This was clearly not the case here, and all these connections resulted in what is known as “pool fragmentation”.

So what’s connection pooling? how does it work, what can cause pool fragmentation and how can we avoid/reduce it?

Read the rest of this entry . . .