Slashdot Mirror
Hope For Fixing Longstanding Linux I/O Wait Bug
DaGoodBoy writes "There has been a long standing performance bug in Linux since 2.6.18 that has been responsible for lagging interactivity and poor system performance across all architectures. It has been notoriously difficult to qualify and isolate, but in the last few days someone has finally gotten a repeatable test case! Turns out the problem may not even be disk related, since the test case triggers the bug only by transferring data either between two processes or threads. The test results are very revealing. The developer ran regressions all the way back to version 2.6.15 that demonstrate this bug has more than doubled the time to run the test in 2.6.28. Many, many people working at improving the desktop performance of Linux will be very happy to see this bug die. I know that I, personally, will find a way to send the guy that found this test case his beverage of choice in thanks. Please spread the word and bring some attention to this issue so we can get it fixed!"
Anyone else notice the article 404ing from the front page? I'd say /. needs to fix some bugs/user errors rather than speak about a Linux IO latency most users don't even notice. Just an observation, and if you can read this, they either fixed it or you doctored up a query string like I did :D.
I'm not sure if this is related, but has anyone else noticed KTorrent can really bog your system down without showing any excessive resource usage in KSysGuard? For all I know, it may be passing information between one thread and another, and it's disk I/O intensive.
Been waiting all of 2 years and change for your precious bug fix, 'ave you? You almost had my eyes tearing up there I tell ya: 25 Year Old BSD Bug.
I think that when it comes to the performance spectrum, Servers would be where this fix is the most needed.
Nope, read the bug. Throughput remains ok, it's the interactivity that suffers.
This is one of those bugs that no Linux developer will admit to until they reckon they have a fix for it. Then we're supposed to be happy, even though people have been complaining about it for years. Oh well, beggars can't be choosers.
I've been ionice'ing my backups and a few other tasks because of this issue, so it'll be nice to get it fixed.
Disk-to-disk operations would then bypass the kernel and asynchronous I/O would consume no primary resources. This was fashionable on some systems (most notably drives that used the IEEE 488 bus) in the 70s and was done to some degree with SCSI, but there's really no excuse for not providing such a capability on any modern drive.
I bought that line, hook line and sinker, in the late 90's with a bunch of IBM 9ES ultra-wide SCSI disks and a good controller.
It never was clear to me that, at any time, Linux was actually telling the drives to copy data directly from one disk to any other without the kernel in the middle.
And now that we live in a world of point-to-point serial buses (SATA, SAS) linking disks to seemingly independent controllers: Is it even theoretically possible anymore?
Kid-proof tablet..
I'm actually pretty sure that I've spotted the results of this in "everyday" use. I've noticed that every once in a long while my hard-drive activity kicks up (it's happened when I'm just scrolling on an already-loaded web page and I'm using absolutely zero swap) and literally everything stops responding for a good 5 seconds. My guess would be that the slocate or "tracker" program spawns off on recently added and removed files, but it's not something I've put a lot of effort into figuring out.
The fscking freezes are in HAL. They have been driving me nuts for more than a year. In my case, the solution is to unplug the CDROM drive.
It's because people don't want to wait for a bugfix for over 2 years. They need fast systems NOW, and when a performance bug which doesn't get fixed can be solved by buying faster hardware, that's what they do.
This is your sig. There are thousands more, but this one is yours.
I am overjoyed that my suspicions have finally been vindicated. I've been working 10+hours a day on linux for the last 13years and you tend to get in tune with your environment (i can still today recite my DOS bootup tune on my XT even though I haven't worked on it for 20 years:-) and some time ago after installing a new flavour of linux I immediately started complaining to fellow workers that something has gone wrong in the kernel but it was not annoying enough to really do something about it; you start living with it. It manifests sometimes when I compile - my system simply locks up for 20-30 seconds which is something I never experienced before. I'd say it happens once out of every 50 compiles of the same program with gcc. During such occurrences, I can't access anything on my desktop which annoyes me cause I typically switch to another kterm session to prepare to run the build whilst compiling (to keep up the productivity and all that). I have also seen strange ratios of i/o to cpu wait in 'top' nowadays but can probably ascribe that to CPU's that just became ridiculously fast and the way top calculates its scores. Nevertheless, I've mumbled over and lambasted i/o wait in Linux ever since a very specific time in the past and even though I haven't noted the exact date, I'm sure its related to this. Anyway, I found this intrigueing enough to create a slashdot account after years to share my joy that the bugs days are hopefully numbered now.
What version do you need to downgrade to? And does downgrading open you up to any security flaws or incompatibility?
http://spamdecoy.net - free throwaway anonymous email - avoid spam!
I'm not sure about anybody else here, but I was surprised to see that they mentioned that this will benefit 'Desktop' users.
They mentioned it because it does hit the desktop: https://bugs.launchpad.net/ubuntu/+source/linux-source-2.6.22/+bug/131094
If you haven't used Linux regularly within the last two years, you probably have not noticed that the system has gotten significantly slower with more recent releases. The probable symptom was discussed here. Many Ubuntu users, including me, have noticed that the latency of desktop operations got significantly larger around the time Gutsy was released, which coincides with the Completely Fair Scheduler and kernel upgrade from 2.6.18.
Since it is most likely a latency issue, the problem is extremely hard to diagnose. Alt-tabbing between programs seem a little slower, keyboard input might lag somewhat. You can't measure desktop latency easily.
Football Odds
It's very easy to trigger, just unrar an iso from a torrent. Regardless of CPU cores, copious amounts of RAM, and no other real system activity, your desktop experience will grind to a miserable halt until the archive process has completed. renicing makes very little difference. Linux has had this problem for years, certainly more than two. Memory suggests it came along with SATA.
Different AC here.
I've already started doing a git bisect on some of the versions. My findings are at home (i'm at work), but there was a difference whereby a newer kernel version had worse performance that an older version.
What tends to happen is that intra-disk and inter-disk transfers are very fast (DMA is on), at the expense of X11 responsiveness. GUI operations (minimise windows, mouse clicks) start to lag. Not sure if this is the same problem as in the article.
I really could do with spending a good few days investigating and bisecting. I hope this is fixed soon.
I think way too many people are blaming their issues on this bug. Some of them may be valid but others probably have something misconfigured or maybe it only affects certain hardware. I don't expereience this bug. My interactivity does not suffer when I do anything I/O intensive.
Time makes more converts than reason
Yes, they do. It's out there in a lot of naming permutations, with a lot of different causes - video, browser, disk, X, general high I/O, etc. I have personally run into this bastard a number of times.
"Genius" is using the 'stable' kernel tree in production systems, when Linus himself recommends vendor kernels for this.
Yep, this is a petty big problem - an easily reproducible one - and it's been around for a really long time. I don't remember when exactly it came about, but I moved from Debian Sid to Ubuntu 7.x about 8 months ago. I didn't have any problem under debian, and I'm uncertain whether the 7.x ubuntus had the problem, but I certainly noticed it in 8.x releases.
I do recall a bit of a somewhat gradual progression of desktop performance decreases, though, going all the way back to the later 2.0 kernels. Back then, the schedulers would all allow an at-the-time relatively slow machine run a fairly bloaty window manager (like E16) responsively while untarring an archive and running a kernel build at the same time - provided there was 100+Mb or so of RAM for the process, of course. Even still, if you were to dip into swap, the UI would remain pretty responsive. Not anymore.
The way things sit now, the Linux I/O scheduler results in desktop performance similar to Windows XP during I/O ops. That is completely unacceptable.
Part of me thinks this is due to a server-centric focus in development (being as the people doing kernel dev largely work for corporations who want server kernels), but I'm not really in the know. If that's the case, we really need to pull one of the old desktop schedulers out of retirement and use that instead of what we've got now, at least for the desktop, and maintain two different-focus schedulers within the kernel instead of just having a couple generally-suited schedulers.
~/ssh slashdot.org ssh: connect to host slashdot.org port 22: too many beers