Messages in this thread | | | Subject | Re: Why the DOS has many ntfs read and write driver,but the linux can't for a long time | From | Lee Revell <> | Date | Mon, 09 Jan 2006 11:56:45 -0500 |
| |
On Mon, 2006-01-09 at 17:53 +0100, Oliver Neukum wrote: > Am Montag, 9. Januar 2006 17:41 schrieb Lee Revell: > > On Mon, 2006-01-09 at 17:29 +0100, Bernd Petrovitsch wrote: > > > On Mon, 2006-01-09 at 11:19 -0500, Lee Revell wrote: > > > > On Mon, 2006-01-09 at 17:14 +0100, Oliver Neukum wrote: > > > > > Am Montag, 9. Januar 2006 17:04 schrieb Lee Revell: > > > > > > On Mon, 2006-01-09 at 17:02 +0100, Oliver Neukum wrote: > > > > > > > Am Montag, 9. Januar 2006 16:15 schrieb Lee Revell: > > > > > > > > On Mon, 2006-01-09 at 15:28 +0100, Oliver Neukum wrote: > > > > > > > > > Am Montag, 9. Januar 2006 15:18 schrieb Robert Hancock: > > > > > > > > > > Yaroslav Rastrigin wrote: > > > > > > > > > > > Well, I could find more or less reasonable explanation of this behaviour - different VM policies of two OSes and > > > > > > > > > > > strangely strong and persistent belief "Free RAM is a wasted RAM" among kernel devs. Free RAM is not a wasted RAM, its a memory waiting to be used ! > > > > > > > > > > > Whenever it is needed by apps I'm launching or working with. > > > > > > > > > > > > > > > > > > > > There is no different VM policy here, Windows behaves quite similarly. > > > > > > > > > > It does not leave memory around unused, it uses it for disk cache. > > > > > > > > > > > > > > > > > > That doesn't mean that the rate of eviction is the same. > > > > > > > > > Is it possible that read-ahead is not aggressive enough? > > > > > > > > > > > > > > > > Enough for what? What is the exact problem you are trying to solve? > > > > > > > > > > > > > > Quicker application startup. > > > > > > > > > > > > Why do you look to the kernel first? The obvious explanation is that > > > > > > Linux desktop apps are more bloated than their Windows counterparts. > > > > > > > > > > It is the most efficient place. An improvement to the kernel will improve > > > > > all starting times. > > > > > > > > I think you'll get at most a 10% or 20% speedup by improving the kernel, > > > > while some of these apps (think Nautilus vs Windows Explorer) will need > > > > to be 1000% faster to seem reasonable to a Windows user. > > > > > > That's easy: Just start nautilus, OOorg, Firefox, a java-vm and > > > GNOME/KDE infrastructure at login time in the background (*eg* and > > > mlockall() the more important ones so that the are surely in RAM) and > > > "starting the app" is only a small program connecting to the respective > > > process to get a fork() there (e.g. like the "-remote" parameter in the > > > Mozilla family). > > > > Have you tried this? I suspect it still takes at least twice as long as > > on windows. > > > > For example on my system there was already a "nautilus" process but > > "Places -> Home Folder" still took ~2 seconds to display anything, and > > ~8 seconds to completely render the window and icons. On Windows this > > takes much less than a second. > > Does the Windows Explorer draw icons based only on name and metadata?
No idea. All i know is it's fast enough to be usable and Nautilus isn't.
Lee
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |