Re: Removing SQLite sessions from the default distribution

On Tuesday, July 1, 2003, at 04:03 PM, Sterling Hughes wrote:

On Tue, 2003-07-01 at 16:20, George Schlossnagle wrote: Not really. The slowdown in locking is still there. Its not an across the board slowdown, but with regards to session handling it is a 1/3 slowdown. What percentage of *total* execution time that is, depends on the script. But when you measure things in a general case, you have to measure the inherent slowdown, not in req/s, but in percentages. Percentage slowdown here is 300%, not 150 req/s.

My point being that 1/3 slowdown seems to be about worst case, given the construction of your benchmark. The test was both designed to exploit lock contention (which does increase in overhead non-linearly under usage due to the queueing issues involved) and to measure only the overhead of the mechanism (by using a script which did nothing else). I'm not disputing your results, just saying you should be more honest about how you interpret them. Oh, and you can't have a '300% slowdown' it's a 66% slowdown. But that's semantics. George