Slashdot Mirror
Why Linux Has Failed on the Desktop
SlinkySausage writes "Linux is burdened with 'enterprise crap' that makes it run poorly on desktop PCs, says kernel developer Con Kolivas. Kolivas recently walked away from years of work on the kernel in despair. APCmag.com has a lengthy interview with Kolivas, who explains what he sees is wrong with Linux from a performance perspective and how Microsoft has succeeded in crushing innovation in personal computers."
Yeah, actually his patches were pretty good. He taught himself C, grokked the kernel coding style, and was a presence on the kernel mailing list. He maintained the -ck set of patches for quite a while, and wrote a couple of new schedulers (staircase, staircase deadline, rotating staircase deadline) based around the concept of fairness.
After quite a bit of discussion, Ingo Molnar produced the CFS (completely fair scheduler) which just recently got merged. The bulk of the new scheduler was written in 62 hours, then finetuned over many weeks on the kernel mailing list. He gave credit to Con for proving the fair scheduler design concept, and for some of the tuning.
A number of people were disappointed by the perceived nepotism, where it appeared that Ingo's got merged because he was in the "in" crowd. I expect this is part of what triggered Con's decision to leave. On the other hand, the two schedulers are very different and it may be that one is really technically better than the other--I haven't compared the two in detail.
Maybe the author doesn't grok the idea of setting the kernel to be responsive for the desktop. It's not rocket science, you know.
Of course not, Microsoft does it for the customer so they don't need to learn how to do it themselves. Would it be so hard for a Linux distro to do so as well when it is doing a "workstation" rather than a "server" install. Some distro ask and have this info regarding intended use.
I think you are exemplifying the "by nerds for nerds" attitude that the author of the article would probably argue is holding back Linux adoption.
His point is that the kernels are optimized for servers. That is, focus on throughput, performance, but not latency or responsiveness. A desktop has the latter two as priorities, while sacrificing the former two. As an example, it doesn't matter if that mpeg4 video I/O eats a little more CPU, as long as other tasks don't interrupt its playback.
This sig does not contain any SCO code.
I want to agree with you, I really do. But my SuSE 10.1 desktop regularly has fits where it becomes completely unuseable - if I can manage to get a shell, I find that the load has spiked to 5-10 (on a single core system) when the system was doing *nothing*. Just this morning, I woke up, poured a bowl of cereal, walked over to it to read some Slashdot over my Cheerios, and found the system thrashing and refusing to come out of screensaver because the load was so high. This happened while I was sleeping. I had to ssh in from my Powerbook to kill off any processes that appeared to be using CPU before the system would respond to the mouse.
Meanwhile at work, we just tossed an Ubuntu server that should have been reasonably swift, but was regularly DOS'ing itself by spiking to loads of 40 or more several times a day under normal use. A load of 40-60, on a single-core machine! We "fixed" it by spending thousands of dollars replacing it with a pair of multicore beast with scads of memory and fast disks, which seems to overpower the problem.
Then there's that server belonging to a client, a RHES 4 system. When I ssh in through a tunnel to update it, it insists on running the update program as an X client for crissakes. Then it tells me to register the system at a URL, but the URL cannot be selected or copied to the clipboard. This is "enterprise" quality software?
Back at work, the dev server is still a RedHat 7.3 clunker. It has a half dozen developers fine-tuning their infinite loops, fork bombs, broken joins, buffer overruns, and spaghetti code, all day long. It simply never crashes or hangs, never gets slow, and never complains about the abuse it receives. It's a rock-solid dream. Except that it's a damn nuisance to update, since it's so old. And it's only hobbyist-quality software, after all, built before RedHat went all enterprise-centric
Posted, with regrets, from my Powerbook. I'm starting to think that software built for the home user is a safer bet than the "enterprise" shite I'm dealing with every day.
First, the Slashdot article is terrible. The article isn't about "why Linux is failing on the desktop", it's about why a kernel developer who was trying to improve scheduling performance quit.
The scheduling issue is interesting. I used to work on mainframe schedulers, I've done real-time work, and I'm familiar with the issue in game implementation, so I know how hard this is. We could do better than what we have now, but not by some magic fix to the scheduler. We have to look at interactivity as a real time problem.
It is, too. Alan Kay used to say that there is no more excuse for a delay between pressing a key on a computer and having something happen than there is on a piano. We haven't been faithful to that, and it subtly drives users nuts.
One useful idea from the real time world is explicit "sporadic scheduling". Some real time operating systems have this. A process can explicitly request that it wants, say, 10ms of CPU time every 100ms. The scheduler must reject that request if the system is overbooked. If it does accept the request, the scheduler has committed that much resource to the process. If the process overruns its time slot, it loses priority and an overrun is tallied.
This is what an audio or video player should be using. This is how you get audio and video that don't pause or skip. For this to work, the player must be able to calculate, for each system it runs on, exactly what resources are needed to play the current content. This may take more analysis and benchmarking than many programmers are used to doing. It's worthwhile to make overruns visible to tools outside the application, so that users can detect broken applications. To a real time programmer, overrunning your time slot means "broken". You have to think that way.
On the interactivity front, it's useful for a thread to be able to request a high priority for a short period after an event, with a priority drop to follow quickly if it keeps the CPU too long. That's how you get the mouse cursor to track reliably. Of course, the thread that handles mouse events has to pass off all the real work to other threads, not stall the thread handling fast events.
It's also probably time to end paging to disk. When it works, paging at best doubles the effective RAM. But paging inherently results in long unexpected delays. If you want interactivity, don't page. Real-time systems don't. Neither do game consoles. RAM is so cheap that it's not worth it. (1GB starts at US$56 today at Crucial.) Paging devices maxed out around 10,000 RPM since the 1960s, and haven't improved much since. Give it up. Today, paging is in practice mostly a means for dealing with memory hogging apps. (Hint: open "about.config" in Firefox and turn off "browser.cache.memory.enable". so it doesn't save screen dumps of each page for faster tab switching.) It's probably time for Linux to not page interactive processes by default.
This implies an operating system that says "no" when you put on too much load, instead of cramming it in and doing it badly. Open too many windows of video, and at some point the player won't open another one. There's nothing wrong with that, but most Linux/Unix apps don't handle resource rejections from the operating system well.
It seems they were sensitive to my complaint and have changed the title of the story at apcmag now. The slashdot title for the interview and their misquoting was... unfortunate.