Autotest White Paper¶
This paper describes the motivation for, and design goals of, the autotest and test.kernel.org projects. Autotest is a framework for fully automated testing, that is designed primarily to test the Linux kernel, though is useful for many other functions too. Test.kernel.org is a framework for communicating, sharing, and analysing test results.
In a traditional corporate systems software development environment, there is normally a large test team responsible for assuring the quality of the final product. Open source projects do not have that luxury, and we need to find another way to run testing. We feel that the only realistic way to achieve the goal is to fully automate the test process, and drastically reduce the need for human staffing. It turns out that this also solves several other critical problems with testing.
- Consistency - it’s much easier to guarantee the tests are run the same way as last time.
- Knowledge capture - the knowledge of how to run testing is not held in one person, but within a system.
- Sharing - you can easily share tests with vendors, partners, and across a wide community.
- Reproducibility - they say 90% of fixing a bug is to get an easily reproducible test case.
Testing is not about running tests … testing is about finding and fixing bugs. We have to:
- Run the tests
- Find a bug
- Classify the bug
- Hand the bugs off to a developer
- Developer investigates bug (cyclical)
- Developer tests some proposed fix (cyclical)
- Fix checked in
- New release issued to test team.
So many test systems I see are oriented only around the first two (or even one!) steps. This is massively inefficient - so often I see developers writing a simple testcase to reproduce what happens in a more complex test, or proprietary application, and then these are thrown away. If we started with open tests that we could freely and easily share, much effort and time would be saved. This is not just about the cost of the people’s time, salaries and machine resources. It’s about the opportunity cost of stalling a release, which is massively greater - these problems are often single-threading the critical path.
We want bug identification, investigation, and fixing to be done earlier in the cycle. This allows multiple debugging efforts to be done in parallel, without affecting others, as well as many other advantages, such as the problem still being fresh in the developers mind, and not interacting with other later changes. This means running tests on multiple codebases (development trees), with high frequency - how can we scale to this? Fully automated testing. Machines are cheap, people are expensive - this is the reality of the modern age. For Linux, the problem is compounded by the staggering diversity of hardware and kernel configurations that we support.
Moreover, a test system is not just about simple functional tests; we should also identify performance regressions and trends, adding statistical analysis. A broad spectrum of tests are necessary – boot testing, regression, function, performance, and stress testing; from disk intensive to compute intensive to network intensive loads. A fully automated test harness also empowers other techniques that are impractical when testing manually, in order to make debugging and problem identification easier, e.g. automated binary chop search amongst thousands of patches to weed out dysfunctional changes.
It’s critical that when operating in an open community, we can share and compare test results - that necessitates consistency of results formats. The easiest way to achieve this is to share one common test harness.
It is critical for any project to maintain a high level of software quality, and consistent interfaces to other software that it uses or uses it. There are several methods for increasing quality, but none of these works in isolation, we need a combination of:
- skilled developers carefully developing high quality code,
- static code analysis,
- regular and rigorous code review,
- functional tests for new features,
- regression testing,
- performance testing, and
- stress testing.
Whilst testing will never catch all bugs, it will improve the overall quality of the finished product. Improved code quality results in a better experience not only for users, but also for developers, allowing them to focus on their own code. Even simple compile errors hinder developers.
In this paper we will look at the problem of automated testing, the current state of it, and our views for its future. Then we will take a case study of the test.kernel.org automated test system. We will examine a key test component, the client harness, in more detail, and describe the Autotest test harness project. Finally we will conclude with our vision of the future and a summary.
It is obvious that testing is critical, what is perhaps not so obvious is the utility of regular testing at all stages of development. There are two main things we’re trying to achieve here, parallelism of work, and catching the bugs as quickly as possible. These are critical as:
- it prevents replication of the bad code into other code bases,
- fewer users are exposed to the bug,
- the code is still fresh in the authors mind,
- the change is less likely to interact with subsequent changes, and
- the code is easy to remove should that be required.
In a perfect world all contributions would be widely tested before being applied; however, as most developers do not have access to a large range of hardware this is impractical. More reasonably we want to ensure that any code change is tested before being introduced into the mainline tree, and fixed or removed before most people will ever see it. In the case of Linux, Andrew Morton’s -mm tree (the de facto development tree) and other subsystem specific trees are good testing grounds for this purpose.
Test early, test often!
The open source development model and Linux in particular introduces some particular challenges. Open-source projects generally suffer from the lack of a mandate to test submissions and the fact that there is no easy funding model for regular testing. Linux is particularly hard hit as it has a constantly high rate of change, compounded with the staggering diversity of the hardware on which it runs. It is completely infeasible to do this kind of testing without extensive automation.
There is hope; machine-power is significantly cheaper than man-power in the general case. Given a large quantity of testers with diverse hardware it should be possible to cover a useful subset of the possible combinations. Linux as a project has plenty of people and hardware; what is needed is a framework to coordinate this effort.
The Testing Problem¶
As we can see from the diagram above Linux’s development model forms an hourglass starting highly distributed, with contributions being concentrated in maintainer trees before merging into the development releases (the -mm tree) and then into mainline itself. It is vital to catch problems here in the neck of the hourglass, before they spread out to the distros – even once a contribution hits mainline it is has not yet reached the general user population, most of whom are running distro kernels which often lag mainline by many months.
In the Linux development model, each actual change is usually small and attribution for each change is known making it easy to track the author once a problem is identified. It is clear that the earlier in the process we can identify there is a problem, the less the impact the change will have, and the more targeted we can be in reporting and fixing the problem.
Whilst contributing untested code is discouraged we cannot expect lone developers to be able to do much more than basic functional testing, they are unlikely to have access to a wide range of systems. As a result, there is an opportunity for others to run a variety of tests on incoming changes before they are widely distributed. Where problems are identified and flagged, the community has been effective at getting the change rejected or corrected.
By making it easier to test code, we can encourage developers to run the tests before ever submitting the patch; currently such early testing is often not extensive or rigorous, where it is performed at all. Much developer effort is being wasted on bugs that are found later in the cycle when it is significantly less efficient to fix them.
The State of the Union¶
It is clear that a significant amount of testing resource is being applied by a variety of parties, however most of the current testing effort goes on after the code has forked from mainline. The distribution vendors test the code that they integrate into their releases, hardware vendors are testing alpha or beta releases of those distros with their hardware. Independent Software Vendors (ISVs) are often even later in the cycle, first testing beta or even after distro release. Whilst integration testing is always valuable, this is far too late to be doing primary testing, and makes it extremely difficult and inefficient to fix problems that are found. Moreover, neither the tests that are run, nor the results of this testing are easily shared and communicated to the wider community.
There is currently a large delay between a mainline kernel releasing and that kernel being accepted and released by the distros, embedded product companies and other derivatives of Linux. If we can improve the code quality of the mainline tree by putting more effort into testing mainline earlier, it seems reasonable to assume that those `customers’ of Linux would update from the mainline tree more often. This will result in less time being wasted porting changes backwards and forwards between releases, and a more efficient and tightly integrated Linux community.
What Should we be Doing?¶
Linux’s constant evolutionary approach to software development fits well with a wide-ranging, high-frequency regression testing regime. The `release early, release often’ development philosophy provides us with a constant stream of test candidates; for example the -git snapshots which are produced twice daily, and Andrew Morton’s collecting of the specialized maintainer trees into a bleeding-edge -mm development tree.
In an ideal world we would be regression testing at least daily snapshots of all development trees, the -mm tree and mainline on all possible combinations of hardware; feeding the results back to the owners of the trees and the authors of the changes. This would enable problems to be identified as early as possible in the concentration process and get the offending change updated or rejected. The test.kernel.org testing project provides a preview of what is possible, providing some limited testing of the mainline and development trees, and is discussed more fully later.
Just running the tests is not sufficient, all this does is produce large swaths of data for humans to wade through; we need to analyse the results to engender meaning, and isolate any problems identified.
Regression tests are relatively easy to analyse, they generate a clean pass or fail; however, even these can fail intermittently. Performance tests are harder to analyse, a result of 10 has no particular meaning without a baseline to compare it against. Moreover, performance tests are not 100% consistent, so taking a single sample is not sufficient, we need to capture a number of runs and do simple statistical analysis on the results in order to determine if any differences are statistically significant or not. It is also critically important to try to distinguish failures of the machine or harness from failures of the code under test.
Case Study: test.kernel.org¶
We have tried to take the first steps towards the automated testing goals we have outlined above with the testing system that generates the test.kernel.org website. Whilst it is still far from what we would like to achieve, it is a good example of what can be produced utilising time on an existing in house system sharing and testing harness and a shared results repository.
New kernel releases are picked up automatically within a few minutes of release, and a predefined set of tests are run across them by a proprietary IBM® system called ABAT, which includes a client harness called autobench. The results of these tests are then collated, and pushed to the TKO server, where they are analysed and the results published on the TKO website.
Whilst all of the test code is not currently open, the results of the testing are, which provides a valuable service to the community, indicating (at least at a gross level) a feel for the viability of that release across a range of existing machines, and the identification of some specific problems. Feedback is in the order of hours from release to results publication.
How it Works¶
The TKO system is architected as show in the figure above. Its is made up of a number of distinct parts, each described below:
The mirror / trigger engine: test execution is keyed from kernel releases; by any -mm tree release (2.6.16-rc1-mm1), git release (2.6.17-rc1-git10), release candidate (2.6.17-rc1), stable release (2.6.16) or stable patch release (18.104.22.168). A simple rsync local mirror is leveraged to obtain these images as soon as they are available. At the completion of the mirroring process any newly downloaded image is identified and those which represent new kernels trigger testing of that image.
Server Job Queues: for each new kernel, a predefined set of test jobs are created in the server job queues. These are interspersed with other user jobs, and are run when time is available on the test machines. IBM’s ABAT server software currently fulfils this function, but a simple queueing system could serve for the needs of this project.
Client Harness: when the test system is available, the control file for that test is passed to the client harness. This is responsible for setting up the machine with appropriate kernel version, running the tests, and pushing the results to a local repository. Currently this function is served by autobench. It is here that our efforts are currently focused with the Autotest client replacement project which we will discuss in detail later.
Results Collation: results from relevant jobs are gathered asynchronously as the tests complete and they are pushed out to test.kernel.org. A reasonably sized subset of the result data is pushed, mostly this involves stripping the kernel binaries and system information dumps.
Results Analysis: once uploaded the results analysis engine runs over all existing jobs and extracts the relevant status; this is then summarised on a per release basis to produce both overall red, amber and green status for each release/machine combination. Performance data is also analysed, in order to produce historical performance graphs for a selection of benchmarks.
Results Publication: results are made available automatically on the TKO web site. However, this is currently a `polled’ model; no automatic action is taken in the face of either test failures or if performance regressions are detected, it relies on developers to monitor the site. These failures should be actively pushed back to the community via an appropriate publication mechanism (such as email, with links back to more detailed data).
Observed problems: When a problem (functional or performance) is observed by a developer monitoring the analysed and published results, this is manually communicated back to the development community (normally via email). This often results in additional patches to test, which can be manually injected into the job queues via a simple script, but currently only by an IBM engineer. These then automatically flow through with the regular releases, right through to publication on the matrix and performance graphs allowing comparison with those releases.
TKO in Action¶
The regular compile and boot testing frequently shakes out bugs as the patch that carried them enters the -mm tree. By testing multiple architectures, physical configurations, and kernel configurations we often catch untested combinations and are able to report them to the patch author. Most often these are compile failures, or boot failures, but several performance regressions have also been identified.
As a direct example, recently the performance of highly parallel workloads dropped off significantly on some types of systems, specifically with the -mm tree. This was clearly indicated by a drop off in the kernbench performance figures. In the graph above we can see the sudden increase in elapsed time to a new plateau with 2.6.14-rc2-mm1. Note the vertical error bars for each data point – doing multiple test runs inside the same job allows us to calculate error margins, and clearly display them.
Once the problem was identified some further analysis narrowed the bug to a small number of scheduler patches which were then also tested; these appear as the blue line (`other’ releases) in the graph. Once the regression was identified the patch owner was then contacted, several iterations of updated fixes were then produced and tested before a corrected patch was applied. This can be seen in the figures for 2.6.16-rc1-mm4.
The key thing to note here is that the regression never made it to the mainline kernel let alone into a released distro kernel; user exposure was prevented. Early testing ensured that the developer was still available and retained context on the change.
The current system is providing regular and useful testing feedback on new releases and providing ongoing trend analysis against historical releases. It is providing the results of this testing in a public framework available to all developers with a reasonable turn round time from release. It is also helping developers by testing on rarer hardware combinations to which they have no access and cannot test.
However, the system is not without its problems. The underlying tests are run on a in-house testing framework (ABAT) which is currently not in the public domain; this prevents easy transport of these tests to other testers. As a result there is only one contributor to the result set at this time, IBM. Whilst the whole stack needs to be made open, we explain in the next section why we have chosen to start first with the client test harness.
The tests themselves are very limited, covering a subset of the kernel. They are run on a small number of machines, each with a few, fixed configurations. There are more tests which should be run but lack of developer input and lack of hardware resources on which to test prevent significant expansion.
The results analysis also does not communicate data back as effectively as it could to the community – problems (especially performance regressions) are not as clearly isolated as they could be, and notification is not as prompt and clear as it could be. More data `folding’ needs to be done as we analyse across a multi-dimensional space of kernel version, kernel configuration, machine type, toolchain, and tests.
As we have seen, any system which will provide the required level of testing needs to form a highly distributed system, and be able to run across a large test system base. This will necessitate a highly flexible client test harness; a key component of such a system. We have used our experiences with the IBM autobench client, and the TKO analysis system to define requirements for such a client. This section will discuss client harnesses in general and lead on to a discussion of the Autotest project’s new test harness.
We chose to attack the problem of the client harness first as it seems to be the most pressing issue. With this solved, we can share not only results, but the tests themselves more easily, and empower a wide range of individuals and corporations to run tests easily, and share the results. By defining a consistent results format, we can enable automated collation and analysis of huge amounts of data.
Requirements / Design Goals¶
A viable client harness must be operable stand-alone or under an external scheduler infrastructure. Corporations already have significant resources invested in bespoke testing harnesses which they are not going to be willing to waste; the client needs to be able to plug into those, and timeshare resources with them. On the other hand, some testers and developers will have a single machine and want something simple they can install and use. This bimodal flexibility is particularly relevant where we want to be able to pass a failing test back to a patch author, and have them reproduce the problem.
The client harness must be modular, with a clean internal infrastructure with simple, well defined APIs. It is critical that there is clear separation between tests, and between tests and the core, such that adding a new test cannot break existing tests.
The client must be simple to use for newcomers, and yet provide a powerful syntax for complex testing if necessary. Tests across multiple machines, rebooting, loops, and parallelism all need to be supported.
We want distributed scalable maintainership, the core being maintained by a core team and the tests by the contributors. It must be able to reuse the effort that has gone into developing existing tests, by providing a simple way to encapsulate them. Whilst open tests are obviously superior, we also need to allow the running of proprietary tests which cannot be contributed to the central repository.
There must be a low knowledge barrier to entry for development, in order to encourage a wide variety of new developers to start contributing. In particular, we desire it to be easy to write new tests and profilers, abstracting the complexity into the core as much as possible.
We require a high level of maintainability. We want a consistent language throughout, one which is powerful and yet easy to understand when returning to the code later, not only by the author, but also by other developers.
The client must be robust, and produce consistent results. Error handling is critical – tests that do not produce reliable results are useless. Developers will never add sufficient error checking into scripts, we must have a system which fails on any error unless you take affirmative action. Where possible it should isolate hardware or harness failures from failures of the code under test; if something goes wrong in initialisation or during a test we need to know and reject that test result.
Finally, we want a consistent results architecture – it is no use to run thousands of tests if we cannot understand or parse the results. On such a scale such analysis must be fully automatable. Any results structure needs to be consistent across tests and across machines, even if the tests are being run by a wide diversity of testers.
What Tests are Needed?¶
As we mentioned previously, the current published automated testing is very limited in its scope. We need very broad testing coverage if we are going to catch a high proportion of problems before they reach the user population, and need those tests to be freely sharable to maximise test coverage.
Most of the current testing is performed in order to verify that the machine and OS stack is fit for a particular workload. The real workload is often difficult to set up, may require proprietary software, and is overly complex and does not give sufficiently consistent reproducible results, so use is made of a simplified simulation of that workload encapsulated within a test. This has the advantage of allowing these simulated workloads to be shared. We need tests in all of the areas below:
Build tests simply check that the kernel will build. Given the massive diversity of different architectures to build for, different configuration options to build for, and different toolchains to build with, this is an extensive problem. We need to check for warnings, as well as errors.
Static verification tests run static analysis across the code with tools like sparse, lint, and the Stanford checker, in the hope of finding bugs in the code without having to actually execute it.
Inbuilt debugging options (e.g. CONFIG_DEBUG_PAGEALLOC, CONFIG_DEBUG_SLAB) and fault insertion routines (e.g. fail every 100th memory allocation, fake a disk error occasionally) offer the opportunity to allow the kernel to test itself. These need to be a separated set of test runs from the normal functional and performance tests, though they may reuse the same tests.
Functional or unit tests are designed to exercise one specific piece of functionality. They are used to test that piece in isolation to ensure it meets some specification for its expected operation. Examples of this kind of test include LTP and Crashme.
Performance tests verify the relative performance of a particular workload on a specific system. They are used to produce comparisons between tests to either identify performance changes, or confirm none is present. Examples of these include: CPU performance with Kernbench and AIM7/reaim; disk performance with bonnie, tbench and iobench; and network performance with netperf.
Stress tests are used to identify system behaviour when pushed to the very limits of its capabilities. For example a kernel compile executed completely in parallel creates a compile process for each file. Examples of this kind of test include kernbench (configured appropriately), and deliberately running under heavy memory pressure such as running with a small physical memory.
Profiling and debugging is another key area. If we can identify a performance regression, or some types of functional regression, it is important for us to be able to gather data about what the system was doing at the time in order to diagnose it. Profilers range from statistical tools like readprofile and lockmeter to monitoring tools like vmstat and sar. Debug tools might range from dumping out small pieces of information to full blown crashdumps.
Existing Client Harnesses¶
There are a number of pre-existing test harnesses in use by testers in the community. Each has its features and problems, we touch on a few of them below.
IBM autobench is a fairly fully featured client harness, it is completely written in a combination of shell and perl. It has support for tests containing kernel builds and system boots. However, error handling is very complex and must be explicitly added in all cases, but does encapsulate the success or failure state of the test. The use of multiple different languages may have been very efficient for the original author, but greatly increases the maintenance overheads. Whilst it does support running multiple tests in parallel, loops within the job control file are not supported nor is any complex `programming’.
OSDL STP The Open Systems Development Lab (OSDL) has the Scalable Test Platform (STP). This is a fully integrated testing environment with both a server harness and client wrapper. The client wrapper here is very simple consisting of a number of shell support functions. Support for reboot is minimal and kernel installation is not part of the client. There is no inbuilt handling of the meaning of results. Error checking is down to the test writer; as this is shell it needs to be explicit else no checking is performed. It can operate in isolation and results are emailable, reboot is currently being added.
LTP (http://ltp.sourceforge.net/) The Linux Test Project is a functional / regression test suite. It contains approximately 2900 small regression tests which are applied to the system running LTP. There is no support for building kernels or booting them, performance testing or profiling. Whilst it contains a lot of useful tests, it is not a general heavy weight testing client.
A number of other testing environments currently exist, most appear to suffer from the same basic issues, they evolved from the simplest possible interface (a script) into a test suite; they were not designed to meet the level of requirements we have identified and specified.
All of those we have reviewed seem to have a number of key failings. Firstly, most lack most lack bottom up error handling. Where support exists it must be handled explicitly, testers never will think of everything. Secondly, most lack consistent machine parsable results. There is often no consistent way to tell if a test passes, let alone get any details from it. Lastly, due to their evolved nature they are not easy to understand nor to maintain. Fortunately it should be reasonably easy to wrap tests such as LTP, or to port tests from STP and autobench.
Autotest - a Powerful Open Client¶
The Autotest open client is an attempt to address the issues we have identified. The aim is to produce a client which is open source, implicitly handles errors, produces consistent results, is easily installable, simple to maintain and runs either standalone or within any server harness.
Autotest is an all new client harness implementation. It is completely written in Python; chosen for a number of reasons, it has a simple, clean and consistent syntax, it is object oriented from inception, and it has very powerful error and exception handling. Whist no language is perfect, it meets the key design goals well, and it is open source and widely supported.
As we have already indicated, there are a number of existing client harnesses; some are even open-source and therefore a possible basis for a new client. Starting from scratch is a bold step, but we believe that the benefits from a designed approach outweigh the effort required initially to get to a workable position. Moreover, much of the existing collection of tests can easily be imported or wrapped.
Another key goal is the portability of the tests and the results; we want to be able to run tests anywhere and to contribute those test results back. The use of a common programming language, one with a strict syntax and semantics should make the harness and its contained tests very portable. Good design of the harness and results specifications should help to maintain portable results.
The autotest Test Harness¶
Autotest utilises an executable control file to represent and drives the users job. This control file is an executable fragment of Python and may contain any valid Python constructs, allowing the simple representation of loops and conditionals. Surrounding this control file is the Autotest harness, which is a set of support functions and classes to simplify execution of tests and allow control over the job.
The key component is the job object which represents the executing job, provides access to the test environment, and provides the framework to the job. It is responsible for the creation of the results directory, for ensuring the job output is recorded, and for any interactions with any server harness. Below is a trivial example of a control file:
job.runtest('test1', 'kernbench', 2, 5)
One key benefit of the use of a real programming language is the ability to use the full range of its control structures in the example below we use an iterator:
for i in range(0, 5): job.runtest('test%d' % i, 'kernbench', 2, 5)
Obviously as we are interested in testing Linux, support for building, installing and booting kernels is key. When using this feature, we need a little added complexity to cope with the interruption to control flow caused by the system reboot. This is handled using a phase stepper which maintains flow across execution interruptions, below is an example of such a job, combining booting with iteration:
def step_init(): step_test(1) def step_test(iteration): if (iteration < 5): job.next_step([step_test, iteration + 1]) print "boot: %d" % iteration kernel = job.distro_kernel() kernel.boot()
Tests are represented by the test object; each test added to Autotest will be a subclass of this. This allows all tests to share behaviour, such as creating a consistent location and layout for the results, and recording the result of the test in a computer readable form. Below is the class definition for the kernbench benchmark. As we can see it is a subclass of test, and as such benefits from its management of the results directory hierarchy.
import test from autotest_utils import * class kernbench(test): def setup(self, iterations = 1, threads = 2 * count_cpus(), kernelver = '/usr/local/src/linux-2.6.14.tar.bz2', config = os.environ['AUTODIRBIN'] + "/tests/kernbench/config"): print "kernbench -j %d -i %d -c %s -k %s" % (threads, iterations, config, kernelver) self.iterations = iterations self.threads = threads self.kernelver = kernelver self.config = config top_dir = job.tmpdir+'/kernbench' kernel = job.kernel(top_dir, kernelver) kernel.config([config]) def execute(self): testkernel.build_timed(threads) # warmup run for i in range(1, iterations+1): testkernel.build_timed(threads, '../log/time.%d' % i) os.chdir(top_dir + '/log') system("grep elapsed time.* > time")
We feel that Autotest is much more powerful and robust design than the other client harnesses available, and will produce more consistent results. Adding tests and profilers is simple, with a low barrier to entry, and they are easy to understand and maintain.
Much of the power and flexibility of Autotest stems from the decision to have a user-defined control file, and for that file to be written in a powerful scripting language. Whilst this was more difficult to implement, the interface the user sees is still simple. If the user wishes to repeat tests, run tests in parallel for stress, or even write a bisection search for a problem inside the control file, that is easy to do.
The Autotest client can be used either as standalone, or easily linked into any scheduling backend, from a simple queueing system to a huge corporate scheduling and allocation engine. This allows us to leverage the resources of larger players, and yet easily allow individual developers to reproduce and debug problems that were found in the lab of a large corporation.
Each test is a self-contained modular package. Users are strongly encouraged to create open-source tests (or wrap existing tests) and contribute those to the main test repository on test.kernel.org (see the Autotest wiki for details). However, private tests and repositories are also allowed, for maximum flexibility. The modularity of the tests means that different maintainers can own and maintain each test, separate from the core harness. We feel this is critical to the flexibility and scalability of the project.
We currently plan to support the Autotest client across the range of architectures and across the main distros. There is no plans to support other operating systems, as it would add unnecessary complexity to the project. The Autotest project is released under the GNU Public License.
We need a broader spectrum of tests added to the Autotest project. Whilst the initial goal is to replace autobench for the published data on test.kernel.org, this is only a first step – there are a much wider range of tests that could and should be run. There is a wide body of tests already available that could be wrapped and corralled under the Autotest client.
We need to encourage multiple different entities to contribute and share testing data for maximum effect. This has been stalled waiting on the Autotest project, which is now nearing release, so that we can have a consistent data format to share and analyse. There will be problems to tackle with quality and consistency of data that comes from a wide range of sources.
Better analysis of the test results is needed. Whilst the simple red/yellow/green grid on test.kernel.org and simple gnuplot graphs are surprisingly effective for so little effort, much more could be done. As we run more tests, it will become increasingly important to summarise and fold the data in different ways in order to make it digestible and useful.
Testing cannot be an island unto itself – not only must we identify problems, we must communicate those problems effectively and efficiently back to the development community, provide them with more information upon request, and be able to help test attempted fixes. We must also track issues identified to closure.
There is great potential to automate beyond just identifying a problem. An intelligent automation system should be able to further narrow down the problem to an individual patch (by bisection search, for example, which is O(log2) number of patches). It could drill down into a problem by running more detailed sets of performance tests, or repeating a failed test several times to see if a failure was intermittent or consistent. Tests could be selected automatically based on the area of code the patch touches, correlated with known code coverage data for particular tests.
We are both kernel developers, who started the both test.kernel.org and Autotest projects out of a frustration with the current tools available for testing, and for fully automated testing in particular. We are now seeing a wider range of individuals and corporations showing interest in both the test.kernel.org and Autotest projects, and have high hopes for their future.
In short we need:
- more automated testing, run at frequent intervals,
- those results need to be published consistently and cohesively,
- to analyse the results carefully,
- better tests, and to share them, and
- a powerful, open source, test harness that is easy to add tests to.
There are several important areas where interested people can help contribute to the project:
- run a diversity of tests across a broad range of hardware,
- contribute those results back to test.kernel.org,
- write new tests and profilers, contribute those back, and
- for the kernel developers … fix the bugs!!!
An intelligent system can not only improve code quality, but also free developers to do more creative work.
We would like to thank OSU for the donation of the server and disk space which supports the test.kernel.org site.
We would like to thank Mel Gorman for his input to and review of drafts of this paper.