For the few other people who are interested, and since it's 3 AM and I
can't sleep, and as long as we're talking long-term directions, let me
suggest a goal to work towards for the next decade or so -- an asset
service which gets faster with use and takes up less absolute space as
more and more people use it. Imagine an algorithm that actually
improves as the number of users N increases, versus one that just gets
worse less rapidly than another as N increases.
Picture that, instead of an asset service that hits a ceiling, like
Second Life, where redoubling the number of users would probably take 8
minutes to retrieve a pair of shoes from inventory.
At a minimum, imagine breaking the barrier from lossless-storage (for
images, say TIFF or PNG) to lossy-storage ( JPG) at user-controlled
acceptable quality levels. With images we're talking reducing some
10 Meg images to 50 K in size, for example. And with successive
resolution, we're talking improving BOTH retrieval speed AND storage
space, with models, like mesh, that are "low", "medium", and "high"
resolutions and using the lowest one that works for, say, a speeding flyby.
But the real power would be in an artificially intelligent system that
did far more than just identifying exact duplicates, but one that hashed
objects into object categories, then conceptualized the object-oriented
definitions and mapped, say, ALL chairs into a "CHAIR" category, where
it only needed to specify the parameters that distinguish THIS instance
of a chair from the base object. And over time, with a Google-like
size memory, it would get better and better at recognizing familiar
categories of objects that people use.
As with the ALICE chatbot strategy, even a dumb parsing algorithm might
handle 80% of what people often do, and 98% of what people almost always
do, which is 95% of the asset server.
Everyone uses "chairs", "vehicles", "walls", "shoes", "walls",
"windows", "doors", etc. and they will predictably continue to do so for
the next decade, so why not get better and better at implementing new
sub-sub-flavors of such common classes? Or, for a creative user, why not
use a "snap-to-grid" type solution? "Hey, are you trying to make a
chair? How about THIS?" (Except please skip the annoying animated
paperclip and puppy part!)
Of course, this would ALSO get vastly better if the entire system,
conceptually, understood hierarchical-objects with unlimited or even 16
sub-groupings, so that the "CAR" category expected that part of most
CAR's would be something like "WHEELS", and there are only a finite
number of types of wheels that would suffice, and for that matter all
four on a vehicle should be instances of each other, as the original
no-duplicates proposal implements.
Then, rather than creating a face with an artists brush, the "rest of
us" could use a police-IDENTIKIT type system to fill-in, top down,
"Chair, office, ..." and in 15 mouseclicks have located in the
marketplace or "created" a new version of "chair" that meets our needs,
and at the same time generated a data-base representation of it as the
12 parameters needed to generate it from known object classes, instead
of storing each and every aspect of each and every chair as if there was
no savings to be had in learning about "chairs" and generalizing.
Wade
On 3/3/12 1:35 AM, M.E. Verhagen wrote:
+1
spitting up the asset tabel is brilliant.
loading an iar or oar of a few 100mb wich was previously saved on the
same grid would than not increase the database with the few 100mb but
just a few 100k :)
Op zaterdag 3 maart 2012 schreef Justin Clark-Casey
([email protected] <mailto:[email protected]>) het volgende:
> Hi folks. As we know, the space required for asset storage in
OpenSimulator grows continuously over time.
>
> I think this is inevitable in a web-like virtual world - distributed
garbage collection is practically impossible. However, the current
OpenSimulator asset service stores much more data than it needs to
since it fails to identify assets that are exact duplicates of each other.
>
> Previous work in places such as OSGrid, which now uses Dave Coyle's
Simple Ruby Asset Service (SRAS) [1], reveals that preventing
duplicate assets has a significant effect on storage requirements (I
can't remember the exact figures but I think that it's >30%).
>
> Therefore, I propose to introduce a new core asset service
(xassetservice) that will implement asset de-duplication via asset
data hashing. This has already been shown to work by SRAS. I regard
this feature as critical for future plans to extend IAR loading and to
improve the 3-months-out-of-the-box OpenSimulator experience. It does
not aim to replace external projects such as SRAS for heavy users.
>
> I already have a working implementation in the xassetservice git
branch (configuration instructions at [2]). This should not be used
in way except for testing - it is still in the prototype stage and can
change at any time. Only a MySQL implementation exists right now.
>
> The plan would be to have xassetservice exist alongside and
independent of the existing asset service. Only one service can be
used at a time and this is determined via config files. After
considerable testing, xassetservice would become the default asset
service for new OpenSimulator installations. The existing asset
service would continue alongside for a very, very, very long time.
>
> Since asset datasets are so large and critical there would be no
automatic migration between assetservice data and xassetservice.
Instead, there would be an external migration tool.
>
> I may also take this opportunity to implement other asset service
features, such as data storage on disk instead of database (possibly
nicking stuff from kcozen's previous patch for this) and maybe
compression (though I'm currently thinking that the cons of this
outweight the pros).
>
> More detail is at [3]. Comments or alternative implementation
suggestions from developers, etc., are welcome.
>
> [1] https://github.com/coyled/sras
> [2]
http://opensimulator.org/wiki/Feature_Proposals/Deduplicating_Asset_Service#Testing
> [3]
http://opensimulator.org/wiki/Feature_Proposals/Deduplicating_Asset_Service
>
> --
> Justin Clark-Casey (justincc)
> http://justincc.org/blog
> http://twitter.com/justincc
> _______________________________________________
> Opensim-dev mailing list
> [email protected] <mailto:[email protected]>
> https://lists.berlios.de/mailman/listinfo/opensim-dev
>
--
Groningen en Hannover Opensims:
secondlife://meverhagen.nl:8002:Hannover ZW/
_______________________________________________
Opensim-dev mailing list
[email protected]
https://lists.berlios.de/mailman/listinfo/opensim-dev
_______________________________________________
Opensim-dev mailing list
[email protected]
https://lists.berlios.de/mailman/listinfo/opensim-dev
