* Geoff Talvola <[EMAIL PROTECTED]> [010604 20:02]:
> At 09:36 PM 6/3/01 -0400, Jay Love wrote:
>
> >We talked about this a while back, and I thought the conclusion was to do
> >it. It's not very complex, anyway. But let's revive the debate....
>
> Last time we went around on this issue, I argued that you're better off
> preallocating all needed threads up front. At least on Windows NT, the
> overhead of a thread that's merely blocking on a Python Queue is basically
> nil. Other OS's may have worse thread implementations though -- for all of
> NT's faults, it seems that threading is one thing it gets right.
Under Linux threads are lightweight in the sense that there is hardly
any work done to create one, Linux uses copy-on-write mechanism so that
only changed pages are copied and only when needed.
That said, it still is a small problem to start of a number of initial
threads upon startup and reuse them when needed.
Probably the best thing is to get this configured by using some form of
a thread allocator class, implement two versions, one that starts
everything on the start and one that changes the number of threads
dynamically.
You have cases for both, the first fits a site with either low traffic
or mostly a constant traffic, the second fits a site that sees large
traffic over a short period of time (Consider the Slashdot Effect).
> I also argued that we'd be better off storing the threads on a stack
> instead of a queue. That way the most recently used threads get reused,
> which may produce slightly better performance. I wonder if there's a
> thread-safe Stack class analogous to the Queue class that comes with Python?
That's most probably a good idea.
--
Baruch Even
http://baruch.ev-en.org/
_______________________________________________
Webware-devel mailing list
[EMAIL PROTECTED]
http://lists.sourceforge.net/lists/listinfo/webware-devel
