[sldev] UDP to TCP/HTTP - performance issues?

[Matthew Dowd]

I've raised this issue on the blog, but thought I would raise it here.

In particular the following text in the recent post mortem blog attracted my attention:

"A server-to-viewer message (related to the mini-map) was updated and changed to move over TCP (reliable, but costly) instead of UDP (unreliable, but cheap and fast). On regions with many avatars, this would cause the simulator to become backed up (storing the “reliability” data) and eventually crash"

There is a lot of chatter in the forums/blogs that server side performance/lag has degraded over recent months. Also people are seeing the client freeze more oftem (see http://jira.secondlife.com/browse/VWR-2051) - typically in such cases the computers cpu is spinning at 100% for some reason.

These seem to coincide with the move of certain protocols from UDP to  TCP/HTTP. I've been suspicious that these might be related for some time (although have no hard evidence). One of the reasons for this, is when I've encountered freezing myself, the CPU runs at 100% even though (a) I can't see any process in the Windows Taskman using up a disproportionate about of CPU (b) I have the affinity of SL locked to just one core, so it shouldn't be able to grab more than 50% CPU. As a result I have been suspicious that it may be a low level driver that is spinning out of control, and one candidate is that somehow SL is overloading the network stack. As the source for the network stack on Windows/Mac/Linux tends to be derived from the same source, it may explan why this problem is not platform specific.

I'm only guessing here, of course, but I do wonder it anyone in LL has done some proper analysis of the performance impacts (under load) of the recent (and planned) protocol changes?

Matthew
_________________________________________________________________
Get free emoticon packs and customisation from Windows Live. 
http://www.pimpmylive.co.uk