[opensource-dev] before opena Jira, APR
Oz Linden (Scott Lawrence)
oz at lindenlab.com
Wed Sep 1 07:47:02 PDT 2010
On 2010-09-01 10:36, Tateru Nino wrote:
> On 2/09/2010 12:20 AM, Oz Linden (Scott Lawrence) wrote:
>> On 2010-09-01 9:46, Francesco Rabbi wrote:
>>> No, this is a client side problem in file handling, not an HTTP
>>> problem... You can parallelize billions of download,
>> Whether or not you _can_, you _shouldn't_. The HTTP spec is quite clear
>> on this point.
> RFC2616 makes for great reading, and the *majority* of it is superbly
> thought through. I spent years with a copy close-to-hand at all times.
>> We'd get much better performance than we're getting now if we fixed the
>> servers to support persistent connections; there's a lot of overhead in
>> setting up a new connection - extra round trips plus TCP slow-start.
> Concur. However, persistent connections (and possibly pipelining) will
> pretty much mean that you'll need to make sure you're maintaining a
> priority-queue of textures to fetch. Otherwise it will *feel* slower to
> the end-user even if it is actually faster in total fetch-and-render
> time. I've been down this road before.
Correct, but that's also something we should improve anyway - doing a
better job of prioritizing which textures are loaded and in what order
could make things _seem_ faster, even if the total time was the same.
Doing both that and the connection handling together would be a big win
(file under "Faster").
More information about the opensource-dev