John & Alain, I am not sure John really intended to cross-post that message, but now that the damage is done, I'll take it one step further: I believe the problem is our sacred open-source cow Linux, or rather that nasty amalgamation of diverse hardware, operating systems, and drivers collectively known as the Linux "workstation".
Both major graphics vendors are reasonably committed to supporting OpenGL on Linux given the limited size of this market as compared to Windows 3D gaming. But in practice, both are overwhelmed by the staggering heterogenaiety that exists. In my view, the task of supporting so many different configurations is fundamentally intractable, and that's why they both fail. Linux is a multitude of moving targets, and as a result, OpenGL on Linux often just doesn't work right -- it may be no one's fault really, but there is simply no one single party who could reasonably be expected to make Linux OpenGL work well. OpenGL is far too complex and demanding for the Linux "state of the art" -- Linux is still too immature a platform to rely upon for this. Case in point: My main development system is a Sun nVidia AMD 64-bit Linux RedHat workstation that STILL, 11 months after purchase, cannot run PyMOL and other 64-bit molecular graphics software without freezing up periodically and exhibiting other major OpenGL glitches. Nothing has worked to solve this problem short of rewriting PyMOL's internal rendering loops in order to pass data at a snail's pace. Who can I blame? nVidia? RedHat? Sun? Linux kernel developers? AMD? John or myself? No one is accountable, and as a company, we're out $5k USD plus many hours of grief for a high-end workstation that just doesn't work. Such stories are far too common with Linux OpenGL. It just doesn't work *every* time like it should. Thus, people cannot expect to rely upon Linux for visualization infrastructure without performing rigourous validation prior to purchase and deployment, and again before all subsequent upgrades. Do people really want to endure such hassles? That is precisely why I am so pumped about Macs -- modern Mac OS X workstations have virtually all the advantages of Linux with almost none of the drawbacks. In my view, any cost premium for Apple is more than made up for in terms of time saved NOT troubleshooting Linux. Apple does the OpenGL validation work for us, and they have a cracker-jack team of experts standing behind their drivers (which they work closely with ATI & nVidia to maintain & optimize). Visualization on Mac OS X just works, and nowadays it works even in windowed stereo 3D! Apple is a responsibility party, and they do care specifically about the needs of science <http://www.apple.com/science/>. While Linux is great for compute clusters and web servers, it just plain-old stinks for visualization. With Mac OS X proven, capable, and mature, can anyone nowadays justify the risks and complications of using Linux for visualization? Apple already has crazy-fast high-end graphics workstations <http://www.apple.com/powermac/graphics.html> and we all expect them to soon release some fast Intel-based portable options as well. To avoid Linux headaches -- my brutually honest advice is to buy Macs instead. It's just that simple. Cheers, Warren PS. Alain, I am sorry the above comments do not help your current situation. For companies with ~50 deployed Linux systems, I really have no other advice other than "good luck" -- we feel your pain. John & I will continue to do our best, but the grief you are experiencing is beyond our ability to fix. With Linux, you just have to keep trying different configurations one after the other until you find one that works. And once you find such a configuration, don't dare change anything (driver, OS, or hardware) until you absolutely have to! -- Warren L. DeLano, Ph.D. Principal Scientist . DeLano Scientific LLC . 400 Oyster Point Blvd., Suite 213 . South San Francisco, CA 94080 USA . Biz:(650)-872-0942 Tech:(650)-872-0834 . Fax:(650)-872-0273 Cell:(650)-346-1154 . mailto:war...@delsci.com > -----Original Message----- > From: John Stone [mailto:jo...@ks.uiuc.edu] > Sent: Thursday, December 08, 2005 11:42 AM > To: alain.dietr...@novartis.com > Cc: v...@ks.uiuc.edu; i...@delsci.com > Subject: Re: ATI as a substitute to nVIDIA > > > Alain, > The brutally honest truth (which I find hard to say on a > public mailing list, but am willing to say in private email) > is that ATI has terrible Linux drivers, and in my experience > thus far, the open source ATI drivers have even more problems > than the commercial ones based on the emails I've gotten from > users of both driver sets. > > Both Warren and I put in a lot of effort making PyMol and VMD > run well on all of the platforms we support, despite the > large number of bugs in the video drivers people are using. > Between ATI and NVidia, ATI is by far the worst offender in > my experience. I've now got a new VMD build that avoids 99% > of the ATI problems, but it is only after having VMD disable > the majority of OpenGL features that trigger problems on > these ATI cards, when it finds that it is running on an ATI card. > Warren and I both have code in PyMol and VMD to sidestep as > many of these problems as we can, but there's a limit to the > amount of time we can spend on this sort of thing, and so > sometimes when ATI and NVidia release a new driver, they > break something new and different, and it takes us a while to > catch up with these workarounds. On Linux, NVidia is still > the best choice out there as far as I can tell, but they too > have problems here and there. On Windows I would actually > recommend going with 3DLabs, as they have the most bug-free > video drivers I've tested to date. > > You'll find that VMD and PyMol are both much more demanding > of the video board that many other applications are. Both > programs use a large percentage of the available OpenGL > functionality, much more than most games or screen savers do. > Their use of much of what OpenGL has to offer makes both > programs very fast and powerful on machines with properly > functioning video drivers, but it is this very same expansive > use of OpenGL that makes our programs vulnerable to a larger > number of bugs in the video drivers. This is indeed the very > reason that I added the OpenGL feature exclusion tests into > VMD to allow the end-user to prevent VMD from using broken > OpenGL features when the vendors release a buggy driver. > > These things go in cycles. At present ATI and NVidia are > both fairly buggy compared to the stability of the old Sun > and SGI workstations. In a few years time I hope that they > will be as stable as the workstations from the commercial > vendors were in years past. If stability doesn't eventually > improve, Warren and I may have to become more conservative in > our use of new OpenGL features as the video driver bugs > constitute a large percentage of the bug reports we both get. > That's certainly true in my case. > > Anyway, if you have ATI cards and you're still having any > trouble with VMD, let me know and I can tell you how to work > around the bugs in their video drivers. > > John > > > On Thu, Dec 08, 2005 at 03:47:43PM +0100, > alain.dietr...@novartis.com wrote: > > > > Dear Warren and John, > > > > We have a substantial park of Linux laptops and workstations (all > > RHELWS 3.x, kernel 2.4.20's), and we are trying to identify some > > alternatives to nVIDIA graphics board (to see what others > are worth, > > to see if some of the freeze problems we have with other > applications > > vanish, etc ... all in all to avoid being bound to them for > the rest of our life). > > > > We have spent some time with an ATI FireGL V5100 (on an HP > xw8200 box) > > and we notice that > > > > ::: with driver 8.19.10 ::: > > VMD 1.8.2 or 1.8.4a22 the drawing of the scene is highly > perturbed for > > certain size of the main window - as if the same molecule was drawn > > several times and stacked on each others at different > viewing angles - . This of course once > > setenv VMD_EXCL_GL_EXTENSIONS GL_ARB_shader_objects is set as > > recommended at : > > http://ati.cchtml.com/show_bug.cgi?id=123 > > http://www.ks.uiuc.edu/Research/vmd/allversions/vmd_bugs.html > > PyMOL 0.97, 0.98rc3, 0.99b27 sometimes crashes at startup > but not always. > > Once launched it is quite stable but we had to set > BackingStore=On in > > X.org to avoid a black main window after > molecular-manipulation (rot.trans.) and "ray". > > > > ::: with driver 8.18.8 ::: > > for both versions of VMD, we still need to set > VMD_EXCL_GL_EXTENSIONS. > > and the problem after main-window-resize stays. > > PyMOL (all three versions) are quite unstable. I have never > seen so many crashes. > > especially after a "show surface", it is relatively easy, > though not > > reproducible at will, to crash it. > > > > > > Given the fact that on our current nVIDIA FX1100, FX3000, > FX3400 (we > > have around fifty of them), once we got the "adequate" > driver (mostly > > of the 6xxx serie) VMD, PyMOL and many other applications > > (Schrodinger/Maestro, Sybyl/Tripos, MOE, Moloc, Wit!P, O, Coot, > > MOLMOL) are running very well and are quite stable, I am > wondering a > > bit about ATI. Note that left apart VMD and PyMOL those 'other > > applications' are stable on the FireGL (either 8.19.10 or > 8.18.8) - which comes rather as a shock to me. Hence this mail. > > > > Do you have good experience with ATI ? Could you recommend > them for production ? > > Do you think it's a dead end and shall we give up ? > > On the PyMOL'wiki I found no mention of ATI - nVIDIA only. > > I read that ATI driver for Linux is (almost) available in > source-code > > form. Which would be a good point in their favor. But after > these troubles, I am wondering a bit. > > > > Any feedback is welcome and appreciated, Best, Alain Dietrich. > > > > -- > NIH Resource for Macromolecular Modeling and Bioinformatics > Beckman Institute for Advanced Science and Technology > University of Illinois, 405 N. Mathews Ave, Urbana, IL 61801 > Email: jo...@ks.uiuc.edu Phone: 217-244-3349 > WWW: http://www.ks.uiuc.edu/~johns/ Fax: 217-244-6078 > > > >