From: jimh@surreal.asd.sgi.com (Jim Helman) Subject: Re: TECH: VR Operating systems and languages Date: Tue, 23 Jun 1992 21:48:52 GMT Organization: Home for Degenerate Physicists In article deering@deering.Eng.Sun.COM (Michael F. Deering) writes: For example, UNIX is not necessarly a single lock kernal. As multi-processor UNIX HW and SW becomes widely available, re-entrent UNIX kernals are quite possible. On the other hand, `real-time'' support is just plain hard in any general purpose system, but the real needs of multimedia will put far more fininacial presure on companies to fix this long before VR is large enough to. Michael is dead on about the utility of multiprocessing for VR work. And a reentrant, highly-preemptable Unix kernel with fine-grained locks is not a mere possibility. IRIX has been this way for several years. There's no doubt that multiprocessing will play a larger role in the future of VR and multimedia. Multiprocessing buys you two very important things: reliable scheduling and high throughput. Scheduling: With all the stuff happening even on a relatively quiescent Unix machine, you'll get hiccups when some daemon or the X server needs to run. For many demanding training applications, hiccups are simply not allowed: the real world does not drop frames. One processor for the VR application and one for Unix fixes this nicely. With the ability to set non-degrading priorities and lock processes to processors, everything can happen when its supposed to. Throughput: CPUs are cheaper than the sort of graphics hardware required for realistic VR whether we're talking about a 486 with DVI boards or a 4/440 VGXT. It's an expensive waste of gfx hardware to have just one CPU which spends half its time doing collision detection or culling and just leaves the gfx hardware idle. In a multiprocessing model, one processor can be dedicated to sending commands down the graphics pipe while the other processors set things up for the next frame doing culling, intersection tests, physics, sound, device I/O, etc. One of the challenges for writing a high performance VR rendering toolkit is to make as much of the multiprocessing (scheduling, shared memory, ring buffers, updating of multibuffered copies) as transparent as possible. rgds, -jim helman jimh@surreal.asd.sgi.com 415/390-1151