Weekly Edition
Return to the Kernel page
| |||||||||||||
Testing for kernel performance regressions
It is not uncommon for software projects — free or otherwise — to include a
set of tests intended to detect regressions before they create problems for
users. The kernel lacks such a set of tests. There are some good reasons
for this; most kernel problems tend to be associated with a specific device
or controller and nobody has anything close to a complete set of relevant
hardware. So the kernel depends heavily on early testers to find
problems. The development process is also, in the form of the stable
trees, designed to collect fixes for problems found after a release and to
get them to users quickly.
Still, there are places where more formalized regression testing could be helpful. Your editor has, over the years, heard a large number of presentations given by large "enterprise" users of Linux. Many of them expressed the same complaint: they upgrade to a new kernel (often skipping several intermediate versions) and find that the performance of their workloads drops considerably. Somewhere over the course of a year or so of kernel development, something got slower and nobody noticed. Finding performance regressions can be hard; they often only show up in workloads that do not exist except behind several layers of obsessive corporate firewalls. But the fact that there is relatively little testing for such regressions going on cannot help. Recently, Mel Gorman ran an extensive set of benchmarks on a set of machines and posted the results. He found some interesting things that tell us about the types of performance problems that future kernel users may encounter. His results include a set of scheduler tests, consisting of the "starve," "hackbench," "pipetest," and "lmbench" benchmarks. On an Intel Core i7-based system, the results were generally quite good; he noted a regression in 3.0 that was subsequently fixed, and a regression in 3.4 that still exists, but, for the most part, the kernel has held up well (and even improved) for this particular set of benchmarks. At least, until one looks at the results for other processors. On a Pentium 4 system, various regressions came in late in the 2.6.x days, and things got a bit worse again through 3.3. On an AMD Phenom II system, numerous regressions have shown up in various 3.x kernels, with the result that performance as a whole is worse than it was back in 2.6.32. Mel has a hypothesis for why things may be happening this way: core kernel developers tend to have access to the newest, fanciest processors and are using those systems for their testing. So the code naturally ends up being optimized for those processors, at the expense of the older systems. Arguably that is exactly what should be happening; kernel developers are working on code to run on tomorrow's systems, so that's where their focus should be. But users may not get flashy new hardware quite so quickly; they would undoubtedly appreciate it if their existing systems did not get slower with newer kernels. He ran the sysbench tool on three different filesystems: ext3, ext4, and xfs. All of them showed some regressions over time, with the 3.1 and 3.2 kernels showing especially bad swapping performance. Thereafter, things started to improve, with the developers' focus on fixing writeback problems almost certainly being a part of that solution. But ext3 is still showing a lot of regressions, while ext4 and xfs have gotten a lot better. The ext3 filesystem is supposed to be in maintenance mode, so it's not surprising that it isn't advancing much. But there are a lot of deployed ext3 systems out there; until their owners feel confident in switching to ext4, it would be good if ext3 performance did not get worse over time. Another test is designed to determine how well the kernel does at satisfying high-order allocation requests (being requests for multiple, physically-contiguous pages). The result here is that the kernel did OK and was steadily getting better—until the 3.4 release. Mel says:
This correlates with the removal of lumpy reclaim which compaction
indirectly depended upon. This strongly indicates that enough
memory is not being reclaimed for compaction to make forward
progress or compaction is being disabled routinely due to failed
attempts at compaction.
On the other hand, the test does well on idle systems, so the anti-fragmentation logic seems to be working as intended. Quite a few other test results have been posted as well; many of them show regressions creeping into the kernel in the last two years or so of development. In a sense, that is a discouraging result; nobody wants to see the performance of the system getting worse over time. On the other hand, identifying a problem is the first step toward fixing it; with specific metrics showing the regressions and when they first showed up, developers should be able to jump in and start fixing things. Then, perhaps, by the time those large users move to newer kernels, these particular problems will have been dealt with. That is an optimistic view, though, that is somewhat belied by the minimal response to most of Mel's results on the mailing lists. One gets the sense that most developers are not paying a lot of attention to these results, but perhaps that is a wrong impression. Possibly developers are far too busy tracking down the causes of the regressions to be chattering on the mailing lists. If so, the results should become apparent in future kernels. Developers can also run these tests themselves; Mel has released the whole set under the name MMTests. If this test suite continues to advance, and if developers actually use it, the kernel should, with any luck at all, see fewer core performance regressions in the future. That should make users of all systems, large or small, happier. (Log in to post comments)
Testing for kernel performance regressions Posted Aug 3, 2012 22:18 UTC (Fri) by rbrito (subscriber, #66188) [Link]
This is great, as all computers that I have access to are what other people would call "old": my fastest machine is like Mel Gorman's hydra computer, with a Phenom II X4 910 (his is a 940).
I will be glad to report whatever results I see with my computers (and I may even throw in some powerpc for comparison).
Testing for kernel performance regressions Posted Aug 3, 2012 22:52 UTC (Fri) by mikov (subscriber, #33179) [Link]
Nobody does regression testing with hardware either. A driver present in the mainline means only that it compiles successfully with that kernel. Nobody has actually tested it. I get scared when I think about that.
Testing for kernel performance regressions Posted Aug 3, 2012 23:32 UTC (Fri) by gregkh (subscriber, #8) [Link]
Then help by testing the hardware you have, that's all any of us can do, right? What else could be expected?
Testing for kernel performance regressions Posted Aug 4, 2012 8:33 UTC (Sat) by fhuberts (subscriber, #64683) [Link]
How about crowd-sourcing that driver testing then?
I'd be willing to donate CPU cycles on most of my machines to test the kernel, drivers, etc. if the results of that would be aggregated somewhere...
Testing for kernel performance regressions Posted Aug 4, 2012 10:11 UTC (Sat) by Cato (subscriber, #7643) [Link]
CPAN Testers already does this for testing CPAN modules and Perl on a vast range of OSes (and hardware) - http://wiki.cpantesters.org/
Organising this for Linux would be harder given a bootable system is required, but it could be done.
Testing for kernel performance regressions Posted Aug 4, 2012 11:49 UTC (Sat) by jnareb (subscriber, #46500) [Link]
> CPAN Testers already does this for testing CPAN modules and Perl on a vast range of OSes (and hardware)
This works so well because by default CPAN client does tests when installing modules, and send those results to CPANtesters. So it is very easy to become CPANtesters contributors.
Perhaps a request at install / upgrade time to perform regression benchmarks of one's system before first run would be a good idea for Linux testers project?
Testing for kernel performance regressions Posted Aug 4, 2012 15:09 UTC (Sat) by Cato (subscriber, #7643) [Link]
That's a good idea - there are already some projects that do functional testing of hardware on new systems, what's missing is the feedback into central test results repository. Though it would need to be a distro-independent test approach so it can be used with the latest kernels as well as distros.
Testing for kernel performance regressions Posted Aug 4, 2012 10:27 UTC (Sat) by siim@p6drad-teel.net (subscriber, #72030) [Link]
i'd also be willing to test on the few machines i have if there was a straightforward way to run the testsuite and to report the results so that they would be visible to interested parties.
Testing for kernel performance regressions Posted Aug 4, 2012 14:42 UTC (Sat) by gregkh (subscriber, #8) [Link]
> How about crowd-sourcing that driver testing then?
That's exactly what we do, and what we expect when we do the -rc releases.
You are running them to test that nothing breaks on your machine, right?
If not, please do so.
Testing for kernel performance regressions Posted Aug 4, 2012 21:34 UTC (Sat) by smoogen (subscriber, #97) [Link]
Well usually the machines that are going to be affected are those in production so running random rc kernels aren't being done on that. My guess is something similar to "Tragedy of the Commons"... people don't test until they have to because they have more important things to do (usually what their boss tells them is important). They then expect that the vendors and coders will do it for them (eg that somewhere there is a huge building filled with racks of machines that vendor X uses every day for every change..) and all for free too.
Testing for kernel performance regressions Posted Aug 6, 2012 10:24 UTC (Mon) by geertj (subscriber, #4116) [Link]
> > How about crowd-sourcing that driver testing then?
> > That's exactly what we do, and what we expect when we do the -rc releases. > You are running them to test that nothing breaks on your machine, right?
Wrong. I am not testing -rc releases, because i have other stuff to do. And i'm not complaining that my hardware doesn't work either, which makes my behavior wholly consistent. Just pre-empting that comment.. :)
There is a lot more that could be done to make this "crowdsourced testing" more effective. Currently it is quite difficult to test out -rc releases. You have to know how to compile a kernel, and how to install and run it in your distribution. Certainly not rocket science, but not easy for the average distro user, which is who you'd need to go after for large-scale outsourced testing.
Just an idea.. What if bootable live test images could be created for -rc releases? Ideally they would need no local storage, but optionally they could use a dedicated partition. The live image could do a whole bunch of tests and send back the results, together with information the hardware the tests ran it. Those could be analyzed for problems. Also the tests could include performance tests.
Every time a new -rc would be released, you'd rebuild the live image, and ask people via G+, Facebook, Twitter, the mailing list, etc, to burn it to a CD and boot their system with it. The CD runs for a few hours overnight, sends back the results, and then says "Thank you, i'm done". I bet you that this could increase your testing base by 10x. Of course you want to be pretty sure that it is safe and put a lot of safeguards in place to make sure it is (which of course doesn't mean you don't need a pretty scary disclaimer before running the CD).
Testing for kernel performance regressions Posted Aug 6, 2012 11:14 UTC (Mon) by niner (subscriber, #26151) [Link]
The idea sounds great. But this way cannot do much more than smoke tests. You can test if the kernel boots on the target's hardware. What you cannot really test is if btrfs will work on the user's drbd device which runs on top of LVM which runs on top of some special hardware RAID controller or such interesting setups. It's at least very difficult to test if suspend/resume works or if the user will experience some performance regression in his favourite game. You cannot test if the external display works automagically when the laptop is put into it's docking station.
In short: it would be a step forward but there's still plenty of stuff which users would have to test manually. But of course: the perfect is the enemy of the good.
Testing for kernel performance regressions Posted Aug 8, 2012 16:33 UTC (Wed) by broonie (subscriber, #7078) [Link]
Lots of the distros do actually have packaged versions of latest -rc or similar kernels available for installation.
Testing for kernel performance regressions Posted Feb 7, 2013 10:52 UTC (Thu) by rbrito (subscriber, #66188) [Link]
> There is a lot more that could be done to make this "crowdsourced testing" more effective. Currently it is quite difficult to test out -rc releases. You have to know how to compile a kernel, and how to install and run it in your distribution. Certainly not rocket science, but not easy for the average distro user, which is who you'd need to go after for large-scale outsourced testing.
Having to compile the kernels is a burden indeed, especially for those with weaker machines.
At least for Debian-based disributions it seems that Canonical provides daily compiled kernels, which is cool to have in mind (I only remembered that when I read your comment):
http://kernel.ubuntu.com/~kernel-ppa/mainline/daily/
Testing for kernel performance regressions Posted Aug 4, 2012 23:11 UTC (Sat) by krakensden (subscriber, #72039) [Link]
Phoronix (insert 5 minute hate) actually already built this. It's just that nobody who does any work on anything it tests ever looks at the results, or talks to them to try and make it more useful for developers.
It's a little sad.
Testing for kernel performance regressions Posted Aug 5, 2012 4:43 UTC (Sun) by dirtyepic (subscriber, #30178) [Link]
Every time I've tried to point out flaws in their compiler benchmarks I've been ignored. I know the signal to noise ratio on their forums is very high but they've continued to make the same mistakes for years now.
Testing for kernel performance regressions Posted Aug 4, 2012 10:30 UTC (Sat) by man_ls (subscriber, #15091) [Link] A technical solution might be to simulate the hardware and test it automatically, perhaps in a special virtual machine. The initial effort would pay off after a few iterations.
Testing for kernel performance regressions Posted Aug 4, 2012 12:08 UTC (Sat) by robert_s (subscriber, #42402) [Link]
Hardware is generally full of subtle bugs and to properly encapsulate them, you basically need a VHDL model of the hardware (that's what VHDL was originally designed for).
Testing for kernel performance regressions Posted Aug 4, 2012 15:21 UTC (Sat) by man_ls (subscriber, #15091) [Link] In this case, anything is better than nothing. The more esoteric specialized hardware will probably harder to emulate; but just testing the most common hardware interfaces would surely save a lot of time and eventually enable devs to change code that may be currently too brittle to touch.Some examples: mount a virtualized SATA disk, format it, create a few files, read them back and check their contents. Mount a virtualized USB disk and do the same. Create several filesystems and stress-test them. Check that the commands issued are in correct order. And so on. This is (obviously) spoken from utter ignorance of kernel internals, just from the point of view of basic software engineering: if it is not tested and verified it is not finished. I am sure kernel devs will know how to implement the idea or ignore it if the effort is not worth it. But for me it would be a fascinating project.
Testing for kernel performance regressions Posted Aug 4, 2012 19:09 UTC (Sat) by robert_s (subscriber, #42402) [Link]
"but just testing the most common hardware interfaces would surely save a lot of time"
But that's the stuff that gets tested anyway, by people. The trouble _is_ with the esoteric hardware.
Testing for kernel performance regressions Posted Aug 4, 2012 22:39 UTC (Sat) by man_ls (subscriber, #15091) [Link] Yes, so it is wasting the time of those people who have to painstakingly compile -rc kernels, load them and check that everything works. Not to speak about performance regressions in the drivers and filesystems, which I imagine must be hard to find and boring work. While a test suite publicly available might be run privately and also on the public git repos after every push. It is called continuous deployment, we do it at my company and it is fun, challenging stuff!
Testing for kernel performance regressions Posted Aug 6, 2012 11:17 UTC (Mon) by niner (subscriber, #26151) [Link]
Testing RC kernels is not that hard. On openSUSE I just added the Kernel_HEAD repo and new kernels are installed with the other package updates. If a kernel breaks my system I can still just boot the old one and report the bug.
Testing for kernel performance regressions Posted Aug 6, 2012 10:27 UTC (Mon) by njd27 (subscriber, #5770) [Link]
Unfortunately simulating VHDL would in most cases be far too slow to perform a realistic test of the kernel.
I actually work on the Linux driver for a particular family of input devices from one manufacturer. Most of the customers are integrating devices which are targeting older kernel versions rather than mainline, so our main focus is there. But we do track patches that are going into mainline and make sure they are sane, and do some occasional testing.
Device driver testing Posted Aug 4, 2012 10:33 UTC (Sat) by ajb (subscriber, #9694) [Link]
In principle, the original hardware companies could do more to help with this. It's difficult to automate tests involving physical devices, and lab space is wanted for the current generation of devices under development. But initial device drivers are often developed on 'models' of the real device before it has come back from manufacturing. The hardware company could set up a regression job, downloading the most recent kernel and booting it on a simulator connected to the model of their device.
In practice, I suspect most hardware companies don't have sufficient incentive to do this, but they might for some products. Some teams also don't really have a model, preferring to develop their device using FPGAs. (I'm not counting RTL models, which are too expensive to run to use for software testing)
Testing for kernel performance regressions Posted Aug 4, 2012 21:47 UTC (Sat) by mikov (subscriber, #33179) [Link]
There are a couple of problems:
- there is no static set of hardware that I use. You can't buy the same hardware even if you wanted to. So, I would have to be testing all the time. Considering that the kernel changes all the time...
- Should I be testing my distro's kernel or the latest mainline? If the latter, why am I testing a kernel I am not going to use for years, if at all?
- where do I even report the bugs? Lkml? Bugzilla? My distro?
In reality, it would be a full time job for a couple of people to deal with all this. Highly paid jobs. Not every business can afford that.
A stable source level API would go a long way towards improving the situation, but that is not likely to happen :-)
Testing for kernel performance regressions Posted Aug 4, 2012 22:11 UTC (Sat) by dlang (✭ supporter ✭, #313) [Link]
> - Should I be testing my distro's kernel or the latest mainline? If the latter, why am I testing a kernel I am not going to use for years, if at all?
you should test the mainline kernel because your distro is going to be based off of the mainline. It's because your distro is based off the mainline that you should test it.
You don't have to test every -rc kernel, but the more testing that you do, the less likely you are to run into problems when you upgrade to the latest release of your distro.
If you test only when your distro upgrades their kernel every 5 years, then finding where in the 5 years of development things went wrong is an impossible task
If you test the released kernels every 3 months, there's only 3 months of work to go through.
If you test each kernel release around the -rc3-5 range, you are even better off as the developers are currently thinking about that release, and looking for problems to fix.
> - where do I even report the bugs? Lkml? Bugzilla? My distro?
That's easy, if you are testing a mainline kernel, LKML is the best place to report bugs.
Testing for kernel performance regressions Posted Aug 3, 2012 23:42 UTC (Fri) by aliguori (subscriber, #30636) [Link]
Is test.kernel.org still down from the break-in?
I had thought the kernel is moving towards autotest for this kind of stuff. autotest is actually pretty capable of harnessing benchmarks like this and collating results.
Testing for kernel performance regressions Posted Aug 4, 2012 23:43 UTC (Sat) by pabs (subscriber, #43278) [Link]
Seems so.
kerneloops.org is also still down :(
Automated testing for kernel performance regressions Posted Aug 4, 2012 10:52 UTC (Sat) by copsewood (subscriber, #199) [Link]
I've developed about 400 regression tests for my own web application because I know from upgrading between production versions with live data well enough why automated testing is needed. But the mind boggles trying to get my head around what's likely to be required in kernelspace - which must be 3 orders of magnitude more complex.
In one sense the current problem resulting from lack of a coherent test facility might be equivalent to the previous practice of trying to manage the kernel patch queue using an email spool without source code revision control. BitKeeper then Git weren't introduced that long ago and their lack must have constrained what Linus could realistically do. I suspect the test problem also inherently likely to get worse until some kind of standardisation and incentive mechanism enables a major distributed community effort to come together into a coherent test facility. Even if it can't cover all automated test requirements, if it could cover enough of these it seems likely (if it's at all feasible) greatly to improve likely quality of released software.
Those with the compute resources likely to be needed to crunch the test software (to the extent hardware emulation is possible), are not always the same as those with the incentive to write the test cases, when it comes to generic as opposed to hardware-specific kernel features. The incentive to contribute hardware emulations would be to get hardware onto a 'platinum level' support list so more purchasers buy it. The incentive to contribute test cases would be so that your tests are automatically run on time and regressions resulting from newer software are automatically reported.
Could the Linux Foundation working with other interested parties attract the resources to fund and develop a cloud type test facility ? This probably wouldn't work for drivers unless either software emulators for the hardware in question exist, or the physical hardware could be installed within the test farm using various standardised protocols allowing for timed tests, automated output comparisons and resets etc. I guess the funding of this would mainly come through hardware manufacturers who want the highest possible level of kernel support.
So if this thought experiment ever leads to feasible development, there seem likely to be 3 main contributors.
a. A vendor independant, trusted and funded community body which runs the main test rig, e.g. Linux Foundation.
b. Contributors of generic test cases for kernel features intended to run on many different types of hardware.
c. Contributors of physical hardware requiring dedicated device drivers, software emulators for that hardware and tests which can run on scaffolded or emulated hardware. (Eventually manufacturers could contractually commit to their customers to supporting physical and or emulated hardware within this facility for a stated period after it ceases to be in production.)
There seem likely to be a few reasons why this isn't feasible, but I can't think of any quite yet.
Testing for kernel performance regressions Posted Aug 4, 2012 22:12 UTC (Sat) by drdabbles (subscriber, #48755) [Link]
One of the major problems with testing the kernel is, as the author notes, the lack of a standard test suite. Picking test suites is like picking religion in most cases- You may be morn into it, you may choose it, or you may not believe in it at all...but everyone has a strong opinion.
The next problem is that there is little incentive for the hardware vendors that contribute directly to the kernel (I'm looking at you, Intel) to also contribute a test suite for their contributions. Moreover, it's quite conceivable that contributing a test suite would allow people to reverse engineer the hardware in question to some extent. I don't see this as a problem, but then again I don't make billions of dollars a year selling chipsets embedded on nearly every device in existence. There may also be regulatory concerns here, much like open-source WiFi drivers had to contend with here in the US several years ago. So, it's a sticky situation.
Additionally, you have the problem of how to actually execute tests. Software bugs aren't always painfully obvious. You don't always get a panic or segfault when a program or driver messes up. In fact, sometimes the program runs perfectly unaware that it has completely ruined data. This problem of subtle bugs can be seen in audio chipset drivers frequently. Sometimes an upgrade causes audio artifacts, sometimes the ports are reversed, and sometimes the power saving mechanisms act wonky. But only after a suspend from a warm reboot where a particular bit in memory wasn't initialized to 0. These things are extremely hard to detect with a test suite, because the suite has no idea if the audio is scratchy or if the port a user expects to be enabled is working properly.
Finally, if you could overcome the issues above, you have the case where suspend/resume/reboot/long run time causes a problem. To test this, the test suite needs complete access to a computer at a very low level. Virtualization will only get you a small portion of the way there. Things like PCI pass through are making this kind of test easier, but that in itself invalidates tests on a very basic hardware access level. This is where the idea of hardware vendors contributing resources to a test farm becomes a great idea. And as a comment above mentioned, the Linux Foundation could create some incentives for this. A platinum list of vendors / devices would be excellent! My organization ONLY uses Linux and *BSD in the field, so having that list to purchase from would be a pretty big win for us.
I think the solution will have to be several layers. A social change will need to be made, where developers don't dread writing test suites for their software. A policy change may be needed such as, "No further major contributions will be accepted without a test suite attached as well". And finally, the technical requirements to actually execute these test suites.
The good news is that booting a kernel with TESTSUITE=yes option could kick the kernel into a mode where only test suites are executed would be pretty easy. The box would never boot the OS, but would sit there running tests for all built (module or in-kernel) components, hammering on hardware to make drivers fail. Passing a further option that points at a storage device could be useful for saving results to a raw piece of hardware in a format that could be uploaded and analyzed easily.
Testing for kernel performance regressions Posted Aug 4, 2012 23:44 UTC (Sat) by pabs (subscriber, #43278) [Link]
Don't pick one test suite, pick all of them!
Testing for kernel performance regressions Posted Aug 4, 2012 23:54 UTC (Sat) by drdabbles (subscriber, #48755) [Link]
At some point, doing that becomes a maintainability nightmare and an impractical approach. Less time should be spent on setting up the test environment than running the tests.
Having to install 2, 3, or 4 test suite packages just to run the tests means nobody will ever actually run them.
Perhaps a solution like GIT- built specifically for the Linux kernel use case, could be helpful.
Testing for kernel performance regressions Posted Aug 5, 2012 2:56 UTC (Sun) by shemminger (subscriber, #5739) [Link]
OSDL tried this, and trust me it failed. The problem is that lots of hardware is expensive to maintain and most tests give no meaningful result. They invested at a least 10 person/years of effort in getting a working test infrastructure and running tests, and the result was only a few bugs that were worth the attention of kernel developers.
Random testing is often better than organized testing! Organized testing works for benchmarks, but the developer in tawain who boots on a new box and reports that the wireless doesn't work is priceless.
Testing for kernel performance regressions Posted Aug 5, 2012 0:03 UTC (Sun) by amworsley (subscriber, #82049) [Link]
From my experience developers are most responsive to comments on the latest changes (which is most to front of mind). I suggest prompt results against each rc release directed to the maintainer of each area would be more developer friendly.
Testing for kernel performance regressions Posted Aug 6, 2012 13:31 UTC (Mon) by cmorgan (subscriber, #71980) [Link]
How about the Phoronix Test suite? http://www.phoronix-test-suite.com/
Phoronix has been doing kernel benchmarking for years now and has pointed out a bunch of kernel releases/distro releases that have performance regressions.
Chris
Testing for kernel performance regressions Posted Aug 6, 2012 22:45 UTC (Mon) by dlang (✭ supporter ✭, #313) [Link]
A large part of the problem is the tests that are part of the suite.
Many of them really don't make sense and some (disk related that I know of) are downright misleading, givng 'better' scores for situations where things are misbehaving.
Many people have tried to point this out and made no progress in getting them changed.
Testing for kernel performance regressions Posted Aug 6, 2012 23:02 UTC (Mon) by Cyberax (✭ supporter ✭, #52523) [Link]
To be fair, that mostly manifests in cases where some filesystem/device ignores 'sync' calls. That usually results in stunningly better performance.
But most other cases are fine. Phoronix usually captures quite real performance regressions.
Testing for kernel performance regressions Posted Aug 7, 2012 13:02 UTC (Tue) by cmorgan (subscriber, #71980) [Link]
The Phoronix test suite is also open source from what the page indicates. Which means that people can provide patches, fork etc.
Just wanted to point out that Phoronix has been a great resource for exactly the kind of benchmarking between kernel releases that the article was referring to, and has been doing so for years.
I wonder if any kernel developers have used the Phoronix results to help target fixes for various regressions.
Testing for kernel performance regressions Posted Aug 8, 2012 21:27 UTC (Wed) by deater (subscriber, #11746) [Link] Part of the problem is that the kernel varies so much from release to release, often in un-git-bisectable ways.I run regression tests of the overhead of the perf_event syscalls. You can see some results here. There are often wild swings in the results of 5-10% from kernel release to release, but it just seems to be "noise". I've tried bisecting but it just gets you nowhere. When I went to the kernel devs they just said it's probably different cache layout and similar affects from unrelated pieces of the code. So sadly regressions are just completely lost in noise unless they are *major* slowdowns.
Testing for kernel performance regressions Posted Aug 8, 2012 23:00 UTC (Wed) by dlang (✭ supporter ✭, #313) [Link]
> So sadly regressions are just completely lost in noise unless they are *major* slowdowns.
However, if there is a series of tests over time, you can show that while there is a 10% noise factor in the tests, over the last X releases, there is a downwards trend or something like that.
|
|||||||||||||