From 0ea5fc66924303d1bf73ba283a383e2aadee02f2 Mon Sep 17 00:00:00 2001 From: neodarz Date: Sat, 11 Aug 2018 20:21:34 +0200 Subject: Initial commit --- pipermail/nel/2001-February/000230.html | 96 +++++++++++++++++++++++++++++++++ 1 file changed, 96 insertions(+) create mode 100644 pipermail/nel/2001-February/000230.html (limited to 'pipermail/nel/2001-February/000230.html') diff --git a/pipermail/nel/2001-February/000230.html b/pipermail/nel/2001-February/000230.html new file mode 100644 index 00000000..d7b56776 --- /dev/null +++ b/pipermail/nel/2001-February/000230.html @@ -0,0 +1,96 @@ + + + + [Nel] Dynamic load balancing? + + + + + + +

[Nel] Dynamic load balancing?

+ Vincent Archer + archer@nevrax.com
+ Tue, 20 Feb 2001 14:20:32 +0100 +

+
+ +
According to Thierry Mallard:
+> On Mon, Feb 19, 2001 at 03:43:59PM -0600, Jared Mark wrote:
+> > Okay, so what you're saying is basically this...
+> > 
+> > By seperating the services, you can have a cluster of machines, each running
+> > different services, and thus, keeping the load down across the board.  You
+> > would have a single system, for example, handling just combat resolution,
+> > and you would have another system that would keep track of the database...
+
+Basically, yes. It may be more complicated than this. For example, we must
+strive to keep only a single service handling items, while we may run
+multiple concurrent instances of AI services to have lots of smart bots.
+
+> We may have to add features concerning high-availability, as this scheme
+> doesn't allow it : if the computing holding the combat resolution fails, the
+> world will be a heaven of peace  ;-)
+
+It's all a compromise. If a geographic server (in a geographic approach)
+fails, suddendly, part of the world becomes unpassable. You can't log in
+if you were in them, you can't enter them, and so on. Different failure
+modes.
+
+> The second point is that if a service grow too much, the load-balancing won't
+> respond to the problem. You mentionned this point in your previous mail I
+> think.
+> 
+> One possible architecture is to write separate process, and give two or three
+> computer the responsiblity to distribute, by monitoring the farm, those
+> process. I dunno if such scheme is possible w.r.t. the current NeL objectives
+
+It is. In fact, a lot of services will be spread across different nodes,
+but that's because these services are 'linear': each agent it manages
+deals chiefly with itself. On the other hand, some are harder to split.
+Item management for example, because of ownership and transfer issues:
+if item A is owner by entity E, and all items of entity E are on server S1,
+what happens when item is given to entity D, whose items are all on the
+server S2. You break locality, which defeats the purpose of functional
+services.
+
+However, for load balancing purpose, one feature of functional split is
+quite simple. If your item server, for example, manages 300,000 currently
+present items, it doesn't matter if everyone currently on-line decides
+to cluster in a single city: there are still 300,000 items to manage, and
+the amount of item interactions haven't changed much.
+
+-- 
+Vincent Archer                                         Email: archer@nevrax.com
+
+Nevrax France.                              Off on the yellow brick road we go!
+
+
+ + + + + +
+

+ -- cgit v1.2.1