User:Zero Linden/Office Hours/2007 Mar 06

From Second Life Wiki
< User:Zero Linden/Office Hours
Revision as of 20:15, 28 March 2007 by Zero Linden (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

Transcript of Zero Linden's office hours:

[13:01] Zha Ewry: Hello Zero!
[13:01] Dale Innis: Hi hi!
[13:01] Jason Ayakashi: hi, zero
[13:01] Zero Linden: hello hello
[13:02] Hiro Market: hi zero
[13:02] Dale Innis apologized for suddenly running off in a few minutes. :)
[13:02] Dale Innis: (apologizes)
[13:02] Zero Linden: Welcome all
[13:02] Zha Ewry: Hey Rex, welcome back
[13:02] Rex Cronon: hello everybody
[13:03] Zero Linden: Everyone had a good weekend?
[13:03] Dale Innis: Even slept some. :)
[13:03] Zero Linden: I spent far too much of mine thinking about cross-protocol endpoints.
[13:03] Dale Innis: lol
[13:03] Zha Ewry: eewe
[13:04] Zha Ewry dislkies m*n problems
[13:04] Zha Ewry: Hello Vitis
[13:04] Rex Cronon: the things just started here right?
[13:04] Vitis Obviate: Heya
[13:04] Rex Cronon: hi
[13:04] Zha Ewry: Zero has even done his "ground rules" yet
[13:04] Zero Linden: Yes - and who would have thought that a simple concept like "the sender of this message" could have such a property?!?!?
[13:04] Zha Ewry would :-)
[13:04] Zero Linden: Yes, things are just started, Rez
[13:05] Zero Linden: wow - big turn out today!
[13:05] flying ball whispers: I am ALIVE!
[13:05] Sifu Moraga: Gotta load my cache....
[13:05] flying ball: , you can't sit on me![cd]
[13:05] flying ball whispers: I am ALIVE!
[13:06] Dale Innis: :) yeah some of us will be running of momentarily.
[13:06] Dale Innis: (off)
[13:06] Sifu Moraga: I'll be falling asleep in a moment
[13:06] Zero Linden: Well - as most of you know, but protocol dictate I say....
[13:06] Zero Linden: This is Zero's Techno-Geek-Fest office hours
[13:06] Zero Linden: And it the transcripts will be posted to the SL wiki
[13:07] Sifu Moraga: Oops, a lot of IBMers here
[13:07] Zero Linden: Unlike last session - I have nothing in particular on the agenda to talk about..... though I'll give a brief update
[13:08] Zha Ewry: Good, then I can pester you about in / out blobs
[13:08] C4 (say /5 boom): Ka-Boom
[13:08] Zero Linden: Last Thursday we were walking about a sim where almost all the sim traffice to the viewer went over LLSD, via the event queue, via a capability
[13:09] Zero Linden: That was a good milestone. We hope this week to have a grid up where the only traffic that is UDP are some messages like ObjectUpdate and the rest
[13:09] Zero Linden: of the servers are all talking LLSD over some TCP based transport, as well as the viewer to sim communications
[13:09] josserand Jacobus: Hi very nice place, Zha ! May I sit down ?
[13:09] Zero Linden: That's all from me
[13:09] Zero Linden: So, I'm happy to talk blobs
[13:10] Zero Linden: For those that don't know, there is a Zero Office Hour Discussion page in the wiki
[13:10] Hiro Market: can i just request the wiki for the last session is updated with the diagrams
[13:10] Zero Linden:
[13:10] Hiro Market: missing atm
[13:10] Zha Ewry: Good. How much pain would it cause to have bigger than 2K objects come in as blobs?
[13:10] Vitis Obviate: ahh..images are up...ty
[13:10] Zero Linden: Hiro - what? has images
[13:10] Sifu Moraga: (and then you'd have to decide how to store them)
[13:10] Zha Ewry: I'm especially keen on the idea, as that gets us out of the LSL data space problems
[13:11] Zha Ewry: As to where Sifu, in inventory, inside the script's prim, is one easy thought
[13:11] Dale Innis: Yeah; get everyone to store the big stuff on their own external srevers! I like that. :)
[13:11] josserand Jacobus: No answer so I sit down
[13:11] Zha Ewry: Because, you don't want them to be live, without some effort on the part of the scripter
[13:11] Sifu Moraga: I currently use notecards for blobs
[13:11] Hiro Market: no, nothing for me, tried opera and ie
[13:12] Zero Linden: So, for others, if you haven't read it, you might want to bop on over to the wiki discussion page and read Zha's proposal
[13:12] Zero Linden: Sifu - right, but notecards arent r/w and aren't likely to be.... ever.
[13:12] Sifu Moraga: yes, notecards are a real pain
[13:12] josserand Jacobus: Zero, what mean LindenVillager please
[13:12] Sifu Moraga: but they were all we could get to work
[13:12] Dale Innis: Hence the desire for more general blobs. :)
[13:12] Zha Ewry: Hmm. Before we dive to far, why aren't notecards writable?
[13:12] Rex Cronon: a link to zha proposal is on you page zero?
[13:12] Sifu Moraga: exactly
[13:12] Zero Linden: Hiro - does this work for you: ??
[13:13] Zha Ewry:
[13:13] Zero Linden:
[13:13] Zha Ewry: If you use mine, scroll down, Zero's is more precise
[13:13] Zero Linden: So
[13:13] Zero Linden: there are two aspects about blobs I'd like to discuss
[13:14] Zero Linden: First - mechanics of what blobs would look like:
[13:15] Zero Linden: Right now, the inventory of an object is essentially a list of assets and access rights information
[13:15] Zero Linden: A binary blob, especially a r/w one could not be stored in the asset server becuase the asset server
[13:15] Zero Linden: (not a single machine, but logically so), was designed, and is cached, on the assumption that
[13:15] Zero Linden: assets are data that never change once written
[13:16] Zha Ewry: I'd assume you'd take it into a native SL type inbound
[13:16] Sifu Moraga: Hmmm, I see the problem
[13:16] Zha Ewry: So, the mime type of the blob would tell the inbound process how to turn it into an SL object, you'ddump it into the asset server
[13:16] Zero Linden: but nonetheless, we'd have to extend the concept of inventory to include objects that aren't assets.
[13:16] Zha Ewry: and the pass the uuid to the invenotry of the prim running the script
[13:16] Sifu Moraga: yes
[13:16] Zero Linden: Not impossible, but definitely different
[13:17] Zero Linden: And once one wants all the inventory like operations to work on them - you see that we'd be effectively trying to
[13:17] Zha Ewry: thus it becomes a fairly normal SL object, before it gets into the asset server
[13:17] Zero Linden: create a sub-class of an abstract class (inventory item) that doesn't currently exist
[13:17] Sifu Moraga: Since an object is local to a sim, could transitory blobs be stored by the sim server?
[13:17] Zero Linden: Sifu - it would have to be -- so
[13:18] Zero Linden: right now, when we move an inventory item (or copy it), there is no data movement at all, becuase the assent doesn't change or move
[13:18] Sifu Moraga: Ko, so the problem is that an object can reference anything that is not in the asset server
[13:18] Zha Ewry: right, you just move a uuid
[13:18] Zero Linden: for a blob, we'd have to hook all those and do copying and/or moving and/or reference counting.... etc...
[13:18] Zero Linden: none of this is impossible, just more difficult and involves a refactoring of a part of the system
[13:19] Zha Ewry: Well, you actually looked at amore general solution... II won't complain if you do that... mind you
[13:19] Zha Ewry: I'd be happy to have a mim driven way of getting blobbed assets into the asset server, and into the inventory of a local object
[13:19] Zero Linden: So now we have to store that data too... and as Sifu said, it would be on the sim... some of the time
[13:19] Sifu Moraga thinks out loud, could an object store a reference to a transient object?
[13:19] Zero Linden: so - when you take an object into inventory (take or take copy), we are basically saving the state of the object (and scripts), and the inventory list, and writing that into a new asset
[13:20] Zha Ewry: I'd asume you'd be reluctant to do to much of that, as you now have state which is in a bad way, if the sim crashes
[13:20] Zero Linden: So in this case, we'd have to bundle all the data of the blob up into the inventory entry and store that too
[13:20] Zha Ewry: right now, assets either are, or don't exist. Not much inbetween
[13:20] Zero Linden: again, not difficult, but more work to do.... and more data to move
[13:20] Vitis Obviate reflects that we should discuss the "why can't we write notecards" problem also - as Zha suggested - before we dive too deep
[13:20] Zha Ewry: In the asset server, with a uuid, or non-existent
[13:21] Zero Linden: Internally, if the object is rezz'd we'd have to have the data hanging off the inventory item... again not hard
[13:21] Zero Linden: and when the sim state saves, we have to write out all that data with objects in the same way as when sent to the asset server during Take
[13:21] Zero Linden: So these changes wouldn't be that hard.. easier than the inventory operation refactoring, actually
[13:22] Zero Linden: But - then there is the issue of the data and the cost of storing it and moving it around
[13:22] Zha Ewry: Vitis, I think you just got a big chunk of your answer to read/write notecard, btw
[13:22] Zha Ewry: Since the same issiues, roughtly obtain
[13:22] Zha Ewry: (lots of state, and the question of the UUIDs of the notecard in each state)
[13:22] Zero Linden: If... imagine, a region with 1,000 objects that used blobs. And, say, 2 blobs per object (pulling numbers out of thin air)
[13:23] Rex Cronon: but what if i am not interested to save the state of the scripts that in an linked set inventory, if i can use scripts to write to notecard, the script can save its onw state
[13:23] Zero Linden: and say 20k of data per blob
[13:23] Zero Linden: that's 20k x 1k x 2 = 40M
[13:23] Hiro Market: somewhere in the forum there's a reference to LL saying they'd open up a database for scripts to read/write data
[13:24] Hiro Market: seperatly to other mechanisms
[13:24] Zero Linden: and, since we store objects in XML -- you'd have to base 64 this when stored
[13:24] Zha Ewry: So, zero,suppose we assume that we use blobs, just to get objects (current set of SL native ones) into/outof the asset server
[13:24] Zero Linden: Since a typical region runs in 200M of RAM, and we try to keep it under 512M
[13:25] Sifu Moraga thinks that the fact that my blobs are already XML doesn't help
[13:25] Zha Ewry: Which scopes things down a lot
[13:25] Zero Linden: you can see that this would start to put a strain on memory
[13:25] Sifu Moraga: But you could still have a local DB, though
[13:25] Zha Ewry: I'd love the more general solution, mind you, but I'm looking for ways for use to unload the storage issue (and the assembly issues) from SL and the scripters
[13:26] Sifu Moraga: not everything hast too be in memory, does it
[13:26] Zero Linden: Hiro - I don't know what forum post you're referring to -- but perhaps you are thinking about silo.php -- which is my script for storing arbitrary amounts of data on your own web server from LSL
[13:26] Zero Linden: ?
[13:26] Zha Ewry: and let us (the developers who own the binary data) get it to the asset server as neatly as possible
[13:26] Zha Ewry: (and the converse, to get it out)
[13:27] Zero Linden: Zha- if your goal is to be able to add textures / sounds / notecards or even (gasp!) scripts to the asset system under LSL control.....
[13:27] Hiro Market: no, not silo, this is a reference to LL agreeing that not writing to notcards was an issue, and supplying a seperate database mechanism accessible by LSL
[13:27] Zero Linden: or similarly, to read that data out....
[13:27] Hiro Market: quite old, looking for it, caught my eye when I saw it
[13:27] Zero Linden: then perhaps we should talk about issuing capabilities to that information so you can just do it directly
[13:27] Zha Ewry: (and follow the DRM model, as we do it...)
[13:27] Sifu Moraga: Certainly that is not what I need blob transfer for
[13:28] Zero Linden: Oh, Hiro, don't know about there - there isn't a current plan internally that I know of
[13:28] Zha Ewry: Capabitlies, with the right strucfture, linked ot scripting, would get us most of the way.
[13:28] Zha Ewry: to
[13:28] Sifu Moraga: you think so, Zha?
[13:28] Zero Linden: Mind you - web server space that runs PHP is so cheep - seems like a much better solution to just have people store their own data in data stores they manage and pay for (and backup!)
[13:29] Zha Ewry: The question is mostly how to get the data from those servers into SL
[13:29] Zero Linden: Rather than the scaling issue of us managing a big writable store for everyeone...
[13:29] Zero Linden: Zha - this brings up my second question... which is - what do you want to do with the blobs
[13:29] Sifu Moraga: Unfortunatly, I'd rather have a standard-like solution like REST or SOAP (gad, ack)
[13:29] Zha Ewry: Now, to an extent, I'd be happy to bypass an asset server, but I'm stuck on how to do that such that it all works when the script moves across sim lines
[13:30] Zha Ewry: REST, not SOAP, please :-) and that's a personal not corporate preference :-)
[13:30] Zero Linden: and perhaps also notecards and sounds
[13:30] Zero Linden: and I'm sensing it is you want to import lots of textures into SL for autmated builds or catalogs or knowledge bases or some such
[13:30] Zero Linden: Ditto here, Zha --- personally, I don't like the taste of SOAP...
[13:30] Sifu Moraga: I know, I know, Zha, but I don't always decide....
[13:30] Zha Ewry: Well also to drive larger data driven scripts.
[13:31] Zha Ewry: We have lots of use cases where we don't (can't) do the compute in SL, nor do you want us to
[13:31] Sifu Moraga: So, what I;m doing with my notecards is to do some light processing on them and then sending them to an instrumented client
[13:31] Zha Ewry: But we need efficient ways of getting the data in, well beyond the 2048 bytes at a time solutions, and with less load on both the scripting workload
[13:31] Zero Linden: Right - but driving data driven scripts is easy - because since the scripts can't compute that much anyway, you just have the scripts request the data in chunks it can handle, which are going to be small
[13:32] Zha Ewry: and the http requests pounding the sims
[13:32] Sifu Moraga: And then, the instrumented client gives a notecard back and gives it to an object via the AV
[13:32] Zero Linden: Really? Surely you can process 2k worth of, say geometry instructions, in LSL in the 1/2 second it takes to fetch the next 2k
[13:33] Zha Ewry: If I want to texture, as I go, it gets worse, and I can drive a nice class five sim down to a time dialation of oh, .4 or ..3 with just geometry and scripts
[13:33] Sifu Moraga: I don't need to do much processing, except a bit of string replacement
[13:33] Zha Ewry: not even loading texutres. Just rezzing prims, and moving them into place in large numbers
[13:33] Zha Ewry: If I want to load textures, as I rez... its gets painful
[13:34] Zero Linden: Right - but I bet the load from the llHTTPRequest isn't really contributing to that problem
[13:34] Sifu Moraga has seen what Zha is up to :-)
[13:34] Zero Linden: We tested doing several 100 llHTTPRequests per second in a sim with no noticable lag at all.
[13:35] Zero Linden: So, for driving scripts, I doubt getting the data in bigger chunks would be any better
[13:35] Rex Cronon: so zha, u want to build something off grid and upload those objects(or linked sets) into a sim?
[13:35] Zero Linden: Now, notecards and textures are two totally other problems
[13:35] Hiro Market: slightly tangental, but has support for procedural textures ever been considered?
[13:35] Zha Ewry: Hmm. Say, a 5500 prim molecular model? Yeah.
[13:35] Sifu Moraga: The problem I have is that I can't transfer the entire notecard in one piece, so I need to do about 4 transfers
[13:36] Zha Ewry: And then stich them together, and then run out of LSL memory
[13:36] Zero Linden: Righy - but doing the, I dunno, 100 2k data requests for the data for that molecule isn't nearly so much the load on the sim as rezzing and moving 'em
[13:36] Sifu Moraga thinks the molecure is awsome
[13:36] Zero Linden: especially since I bet each of those has a script!
[13:36] Zha Ewry: One problem I see, is we end up with lots of scripts to deal with dataspace
[13:36] Zha Ewry: Not for long. We kill 'em as fast as we can.
[13:37] Zha Ewry: If I leave 5500 scripts listening, the sim is an unhappy place for 2 agents
[13:37] Zero Linden: none the less, the cost of starting a script far outweights the cost of doing one llHTTPRequest
[13:37] Zero Linden: Let's talk notecards -
[13:37] Zero Linden: Sifu - you seem to be wanting to do two things with them....
[13:37] Zha Ewry: Hmm. I'm not surprised. Scripts require lots of context
[13:37] Zero Linden: One is transmit coded data to a specially instrumented browser
[13:38] Sifu Moraga nods
[13:38] Vitis Obviate would use that also
[13:38] Zero Linden: You could, of course, use specially coded IM messages -- (what is that crypto product that works over AIM???)
[13:38] Zha Ewry as well
[13:38] Zero Linden: but i imagine that is slow....
[13:39] Zero Linden: So - why not put the data via llHTTPRequest onto your server, and have the client retrieve it?
[13:39] Sifu Moraga: We thought of that, but still have to get it out to a web service at one point
[13:39] Zero Linden: granted, you have to take up the store of it
[13:39] Zero Linden: right - but the instrumented data is for the viewer, not the human, right?
[13:39] Sifu Moraga: that too
[13:39] Zero Linden: is it larger than 2k?
[13:39] Sifu Moraga: It also needs a bit or processing in-world
[13:39] Sifu Moraga: The data is for a cpu, its all XML, but unreadable
[13:40] Sifu Moraga: it is about 16kb in it's least cryptographicaaly secure form
[13:40] Zha Ewry: A fair bit of crypto/cert stuff ends up bigger than 2048...
[13:40] Zero Linden: We've had requests to have a version fo llHTTPRequest that is issued by the viewer, not the sim
[13:40] Sifu Moraga: easily
[13:40] Zha Ewry: Ah. There's a nice hard number
[13:41] Sifu Moraga: not the same problem, but still of interest for other stuff
[13:41] Zero Linden: so a script on the sim makes the request, the request is sent to the viewer, the viewer makes the call, and teh results piped back up to the
[13:41] Zha Ewry: Thus inside the client's firewall, and access to the local machine?
[13:41] Zero Linden: sim
[13:41] Sifu Moraga: yup
[13:41] Zero Linden: Right - the desire is to allow HUD objects that control things like MP3 players, and other, er, "add-ons"
[13:41] Sifu Moraga: more or less, though the firewall is incidental to what we are doing
[13:41] Zha Ewry: That helps although if its 2048, only a little, as you're still in the data stitching game
[13:42] Zero Linden: This strikes me as more of what you are after in this case, Sifu?
[13:42] Sifu Moraga: important is that we have two completely separate zone we need to access
[13:42] Zero Linden: Well, Zha, if the data has to go through LSL, while we can relax the size a bit (we intend to make it configuratble at some point)
[13:42] Zero Linden: it still won't be useful beyond, say 6k
[13:43] Zha Ewry: Right. Thus the desire to bypass LSL
[13:43] Zero Linden: since LSL execution space is going to be limited to 16k for quite some time
[13:43] Sifu Moraga: 16 kb is the minimum
[13:43] Zha Ewry: I'm looking at how I can move bigger chunks without having them have to land in LSL at all
[13:43] Zero Linden: But - the only thing I think I hear that you'd entirely bypass LSL for is: Textures, and Notecards for users to read
[13:43] Sifu Moraga: It did occure that if we could just send the notecard directly from llHTTPRequest....
[13:43] Zha Ewry: Well, notecards for programs as well
[13:43] Vitis Obviate: yes
[13:43] Sifu Moraga: ...that would bypass the LSL memory restrictions
[13:44] Zero Linden: Right - but where? If not a texture or a human readable notecard, then all other destinations in SL require LSL, yes?
[13:44] Sifu Moraga: I do think the use ofr notecards to be just wrong, somehow
[13:44] Zha Ewry: I'm happy to loop across a long notecard, one line at a time, never getting big chunks into the LSL data space
[13:44] Vitis Obviate nods
[13:44] Zero Linden: Then why not read the data from your web server one line at a time? It is almost as fast, if not faster!
[13:44] Sifu Moraga: umm, you'l be happy until you have tried it like me
[13:45] Zha Ewry: I'm muchg less happoy trying to coordinate that across http gets with all the logic to deal with dropped /oiur of order stuff
[13:45] Sifu Moraga: I get some delay that the users don't like
[13:45] Sifu Moraga: yes, that is worse
[13:45] Zero Linden: Indeed, the notary reads notecards one line at a time (needs to for hashing) and it is a bear
[13:45] Zha Ewry: I also see about 50% of my code space go into the parse/manage packet order problem
[13:46] Zero Linden: Really? Why not just issue your HTTP requests serially, like you'd do for notecards?
[13:46] Sifu Moraga: Also, I can't get llHTTPrequest to add chunking headers
[13:46] Zha Ewry: I understand the 16K problem, so I'm looking for reaosnale ways to bypass it
[13:46] Zha Ewry: I can issue them serially, but I have to have chunking/retry logic and I do end up exercising ti
[13:46] Zha Ewry: it
[13:47] Zero Linden: By the way - I'm not trying to be resistant - just trying to really get at what the user needs are here -- alas, with our highly constrained system, and our opaque implementation, it is hard to know what the crux of a user problem is
[13:47] Sifu Moraga totally understands
[13:47] Zha Ewry: Yes, I get that And having the discusssion is the right way to do it
[13:47] Zha Ewry: I'm totally happy witrh this process, at the moemnt
[13:48] Zero Linden: ah- I think I'm now seeing the bigger issue - here
[13:48] Zha Ewry: Point a number of people at the problem, and get to a good technical answer works for me
[13:48] Zero Linden: I'm wondering then if instead we supported byte-ranges,
[13:48] Sifu Moraga: If I was travelling to SF any time soon, I'd love to have a RL discussion of it
[13:48] Sifu Moraga: (unfrotunatly, I'm not)
[13:48] Zero Linden: and since our requests are squid cached
[13:48] Zero Linden: if that would work - because squid requests the whole thing when you do a byte range request
[13:49] Zero Linden: so once the frist range worked
[13:49] Zero Linden: you'd be getting the rest out of the cache
[13:49] Zha Ewry: Oh, allow us to load a bigger chunk into the squid, and then grab it 2048 at a time?
[13:49] Sifu Moraga thinks
[13:49] Zero Linden: Exactly- if everyone played right by the headers and all....
[13:50] Sifu Moraga doesn't get it
[13:50] Zero Linden: If you do a GET to say
[13:50] Zero Linden: but include a byte range header asking for bytes 0-2047
[13:50] Zha Ewry willing to be a very obedient header wreiter
[13:50] Zero Linden: Squid cache will instead ask for the whole thing
[13:50] Zero Linden: and cache the whole resource at that URL
[13:50] Zero Linden: and deliver only the requested bytes
[13:50] Zero Linden: when you then make a second request to the same URL
[13:51] Zero Linden: but with byte range 2048-4095
[13:51] Zha Ewry: with the byterange...
[13:51] Zha Ewry: Yes..
[13:51] Zero Linden: squid just answers out of the cache
[13:51] Vitis Obviate will use whatever incontonation is required
[13:51] Zha Ewry: How big a chunk could the squid handle?
[13:51] Zero Linden: Now, mind you, it is all REST and stateless... so
[13:51] Zha Ewry likes REST and stateless...
[13:51] Zero Linden: in theory, you could have squid drop the resource in the middle and have to refetch
[13:51] Vitis Obviate: fine with me
[13:51] Sifu Moraga thinks REST is good
[13:51] Zha Ewry: Yes, just delay on random fetches if it does
[13:52] Zero Linden: Squid? Oh, I don't know what our maximum cached object size is set to....
[13:52] Zha Ewry: as long as the URL is idempotent, we're ok
[13:52] Zha Ewry smiles
[13:52] Zero Linden: right now it is probably pretty large... 1M or so?
[13:52] Zero Linden: Exactly, Zha
[13:52] Zha Ewry: I can deal with that :-)
[13:52] Zha Ewry: Doesn't help with textures or sounds, but... it would help a lot.
[13:52] Zha Ewry: Be harder to do in the other direction, I'd imagine
[13:53] Zero Linden: Yes - but there there is the much more natural solution - just let you texture an object with a URL to your own image on your own server...
[13:53] Zero Linden: Not that we are doing that right now....
[13:53] Zha Ewry: Yes.
[13:53] Zero Linden: We do have to be careful about opening the flood gates to textures
[13:53] Hiro Market: like the sound of that
[13:54] Zha Ewry: That'd be fine, within reason. How exactly we get the proggramitrc control is interesting, but I can see ways.
[13:54] Zha Ewry: Yes, I'm wiling to hear lots of stories about throttle on textures
[13:54] Zero Linden: Obviously if everyone just uploaded a texture for every surface they needed, under LSL command... we'd see things like
[13:54] Hiro Market: from my perspective one of the biggest restrictions on LSL is not being able to programatically manipulate textures
[13:54] Zero Linden: billboards that uploaded new textures every 15 seconds
[13:54] Rex Cronon: it would be ok to me if instead of html i can use svg
[13:54] Hiro Market: at pixel level
[13:54] Zero Linden: which would be cool from the SL experience point of view
[13:54] Zha Ewry: Yes... throttleing is an issue, of course.
[13:54] Zero Linden: but would really kill the asset server
[13:54] Zha Ewry: Althoug.. if you can have "transient" textures
[13:55] Sifu Moraga: It does sound like you will need the concept of transient blobs stored by the sims
[13:55] Zha Ewry: locally cached, I can come up with a story that only pounds the sim
[13:55] Zha Ewry: And not even that bad.
[13:55] Zero Linden: Hiro - you aren't likely to see things like paramatic textures or pixel level texture control from LSL
[13:55] Zha Ewry: The sim has to fetch/cache the tyextures today
[13:55] Zero Linden: These are memory and process intensive operations that the sim doesn't have the resoruces to do
[13:55] Zha Ewry: Nor should it
[13:55] Zero Linden: what is more likely, what scales much better, is
[13:55] Zha Ewry: Keep as much rendering on the client as possible
[13:55] Hiro Market: again I wonder about procedural textures on the client with shader code on server
[13:56] Zha Ewry nods at Hiro
[13:56] Zero Linden: do the pixel level processing on your own computer or server - and then have a way to put that texture on an object
[13:56] Zero Linden: The servers, the sims, have no graphics hardware at all... so no OpenGL going on there
[13:56] Zha Ewry: I think mosgt of us who have loooked at the scale issues want a smuch rendering as possible on the client
[13:56] Zha Ewry: I don't want the SIMs even thinking about rendering
[13:56] Hiro Market: don't need it, just the shader code on server, rendered on client
[13:57] Zero Linden: don't expect the sims to be doing graphic manipulation or rendering (which is a hint, by the way, as to why the map is such a pain point for us!)
[13:57] Zha Ewry: Among other horrors, you end up with the sim needing to know things like the resolutoin of every client they are talking to
[13:57] Zero Linden: Ah - but Hiro, can you imagine what it would be like if an object could down load arbitrary shader code to your viewer?!?!?
[13:57] Hiro Market: but procedural shader code is simply code, no rendering on server required at all
[13:57] Zero Linden: Talk about griefing!
[13:58] Hiro Market: hmm, point taken, but the advantages must be worth exploring surerly?
[13:58] Zha Ewry: He. Just because sooner or later we'd have all the pron spam in the shader :-)
[13:58] Zero Linden: Alas, OpenGL doesn't have virtualization or sandboxing facilities to make that reasonable
[13:58] Zero Linden: Nor does anyone have shader verifiers as far as I know....
[13:58] Zero Linden: But - one could imagine in
[13:58] Zero Linden: the future
[13:58] Sifu Moraga thinks there might be a case for somre rendering on the server, but it would probably need a different architecture
[13:59] Hiro Market: 3dlabs have one
[13:59] Zha Ewry shudders at what you could get the shader to do.
[13:59] Hiro Market: a verifier that is
[13:59] Zero Linden: private sims that said "If you come here, I will use spiffy special purpose shaders.... allow?"
[13:59] Rex Cronon: talking about griefing, what exactly does the physics engine do to attachements, when the one that wears them is blitzed?
[13:59] Zha Ewry: Hm. That'd be interesting
[13:59] Zha Ewry: In general, the question of private sim behavioir is interesting
[13:59] Zero Linden: Hiro - curious - good to know - though again I wonder if it worries about hostile environments or just makeing sure things won't crash...
[14:00] Sifu Moraga: My cranial CPU is shutting down, and my Powerbook is out of juice, so I'm off early
[14:00] Sifu Moraga: cu all
[14:00] Zero Linden: Yes, it is.... short term, once we start letting estate owners control upgrade times.... we are going to have to deal with
[14:01] Zero Linden: viewers and regions with different feature sets....
[14:01] Zha Ewry: Yes
[14:01] Zero Linden: Well, Sifu - actually it is 2pm so I'm signing off soon too
[14:01] Zero Linden: Shorter term -
[14:01] Zha Ewry: we talked about tadeoffs, between number of scirpts and data spaces and such
[14:01] Zha Ewry: last week.
[14:01] Zha Ewry: So lots of things I'd love to be able to contrrol on the sim :-)
[14:02] Zero Linden: one could easily imagine custome viewers that have special shaders in 'em and then some way for objects
[14:02] Zero Linden: to invoke that shader behavior if running on viewer that has 'em
[14:02] Zha Ewry: Wow. This was incredbly useful Zero. Hope it was for you. I'll try and update my blob comments a bit. and think about some of your comments
[14:02] Zero Linden: Zha - yes - though realize, like all server tuning, the more parameters you have, the better performance you can achieve, but the more compicated it is for people to manage it
[14:03] Zha Ewry: Hey I race saliboats. More strings, more fast. or more slow.
[14:03] Zha Ewry: so, yes.
[14:03] Zero Linden: and - one has to realize that you can only do so much, even with so many knobs
[14:03] Zero Linden: (thinks of tuning squid, or apache, or mysql....)
[14:03] Zha Ewry: Yep. 1 cpu, 512M of memory, etc.
[14:03] Zha Ewry: But, there are cases where we can play those off well
[14:04] Zero Linden: It is great - I'm very happy with my office hours experiences - and have been talking up how good they are inside of LInden
[14:04] Zha Ewry: Thanks again Zero, as always. Really focused discussion today
[14:04] Zero Linden: Thanks all for coming
[14:04] Zero Linden: See you next time.
[14:04] Kooky Jetaime: Thank you Zero
[14:04] Vitis Obviate: very valuable info - ty
[14:04] Drewan Keats: Thanks Zero.
[14:04] Hiro Market: ty zero, aoppreciated
[14:04] Zha Ewry: Thanks everyone, actually, very good discussion today