Performance counters are a piece of cake to use from managed code (especially compared to the equivalent unmanaged code).

They also have the advantage that you can do it remotely (at least assume both machines are on the same domain, or belong to trusted domains). That's probably not that likely in a typical client/server environment though.

I think the overhead is fairly limited. Obviously, things like memory or CPU usage are already tracked by the system anyway. I have noticed though, that querying this kind of information from a remote server does take some time. I have build a web application that shows statistics for about ten servers, including memory usage, number of anonymous users per application, SQL server connections etc. It may have something to do with the connection (and the related security processes) to the server itself, rather then the collection of the requested information. The system doesn't keep the connection alive (like an open pipeline or at least a connection pool), but it reconnects every time the users ask for a refresh of information. I guess these are issues to take into account.

I do wonder though, what the value in polling things like CPU time every 5 seconds is. Depending on the applications running on the server, these statistics can fluctuate heavily even in short periods of time. It's not impossible to imagine the polling of data itself has a significant influence on these statistics.