It really kind of depends on your business and your systems. For example, I work for an insurance company. We don't do umpty-gazillion transactions per/second, so measuring just transactions per second wouldn't supply us with much information. Instead, more often than not, we look at query execution time. But even that is not enough of a measure. Instead of finding a single number and deciding that's your point of entry, I'd suggest just a few measures. Base everything on waits and queues. Disk Queue Length, Processor Queue Length and various system waits. Gather those metrics and you'll have a good idea of the performance of the system. Anything else is just a symptom. Transactions/sec went down... why? We have an increased disk queue length. Uh, oh, IO issues. See what I mean?