MMTests is a configurable test suite that runs performance tests against arbitrary workloads. This is not the only test framework but care is taken to make sure the test configurations are accurate, representative and reproducible. Reporting and analysis is common across all benchmarks. Support exists for gathering additional telemetry while tests are running and hooks exist for more detailed tracing using ftrace or perf.
The top-level directory has a single driver script called run-mmtests.sh
which reads a config file that describes how the benchmarks should be
configured and executed. In some cases, the same benchmarking tool may
be used with different configurations that stresses the scenario.
A test run can have any name. A common use case is simply to compare kernel versions but it can be anything —different compiler, different userspace package, different benchmark configuration etc.
Monitors can be optionally configured, but care should be taken as there is a possibility that they introduce overhead of their own. Hence, for some performance sensitive tests it is preferable to have no monitoring.
Many of the tests download external benchmarks. An attempt will be made
to download from a mirror if it exists. To get an idea where the mirror
should be located, grep for MIRROR_LOCATION=
in shellpacks/
.
A basic invocation of the suite is
$ ./bin/autogen-configs
$ ./run-mmtests.sh --no-monitor --config configs/config-pagealloc-performance 5.8-vanilla
$ ./run-mmtests.sh --no-monitor --config configs/config-pagealloc-performance 5.9-vanilla
$ cd work/log
$ ../../compare-kernels.sh
$ mkdir /tmp/html/
$ ../../compare-kernels.sh --format html --output-dir /tmp/html > /tmp/html/index.html
The first step is optional. Some configurations are auto-generated from a template, particularly the filesystem-specific ones.
Note that List::BinarySearch
and maybe even Math::Gradient
may
need to be installed from CPAN for the reporting to
work. Similarly, R
should be installed if
attempting to highlight whether performance differences are statistically
relevant.
A tutorial with some more details and the full output of each step is available here:
All available configurations are stored in configs/
.
For example config-pagealloc-performance
can be used to run tests that
may be able to identify performance regressions or gains in the page allocator.
Similarly there are network, disk and scheduler configs.
The config file can take many options, in the form of export
-ed
variables. there is a functional sample config file available in
config
.
Some options are universal, others are specific to the test. Some of the universal ones are:
MMTESTS
: A list of what tests will be run.AUTO_PACKAGE_INSTALL
: Whether packages necessary for building or running benchmarks should be automatically installed, without asking any confirmation (takes ayes
or ano
). Creating a file called.mmtests-auto-package-install
and putting it in/
would be equivalent of having this set toyes
.MMTESTS_NUMA_POLICY
: Whethernumad
ornumactl
should be used for deciding (typically, for restricting) on what CPUs and/or NUMA nodes the benchmark will run. It accepts several values.none
,numad
orinterleave
, are the simplest, but the following ones can also be used:fullbind_single_instance_node
fullbind_single_instance_cpu
membind_single_instance_node
cpubind_single_instance_node
membind_single_instance_node
membind_single_instance_node
cpubind_largest_nonnode0_memory
, in which case,MMTESTS_NODE_ID
should also be definedcpubind_node_nrcpus
, in which caseMMTESTS_NUMA_NODE_NRCPUS
should also be defined. Ifnone
is used or the option is not present, nothing is done in terms of NUMA pinning of the benchmarks.
MMTESTS_TUNED_PROFILE
: Whether or not the tuned tool should be used and, if yes, with which profile. In fact, the option takes the name of the desired profile (which should be present in the system). If this is definedtuned
is started and stopped around the execution of the benchmarks.SWAP_CONFIGURATION
,SWAP_PARTITIONS
,SWAP_SWAPFILE_SIZEMB
: It's possible to use a different swap configuration than what is provided by default.TESTDISK_RAID_DEVICES
,TESTDISK_RAID_MD_DEVICE
,TESTDISK_RAID_OFFSET
,TESTDISK_RAID_SIZE
,TESTDISK_RAID_TYPE
: If the target machine has partitions suitable for configuring RAID, they can be specified here. This RAID partition is then used for all the tests.TESTDISK_PARTITION
: Use this partition for all tests.TESTDISK_FILESYSTEM
,TESTDISK_MKFS_PARAM
,TESTDISK_MOUNT_ARGS
: The filesystem,mkfs
parameters and mount arguments for the test partitions.TESTDISK_DIR
: A directory passed to the test. If not set, defaults toSHELLPACK_TEMP
. The directory is supposed to contain a precreated environment (eg. a specifically created filesystem mounted with desired mount options).STORAGE_CACHE_TYPE
,STORAGE_CACHING_DEVICE
,STORAGE_BACKING_DEVICE
: It's also possible to use storage caching.STORAGE_CACHE_TYPE
is either "dm-cache" or "bcache". The devices specified withSTORAGE_CACHING_DEVICE
andSTORAGE_BACKING_DEVICE
are used to create the cache device which then is used for all the tests.
It is possible to retrieve information about the characteristics of the system where the benchmarks will be running, and use them inside a config file.
For instance:
MEMTOTAL_BYTES
: Tells how much memory there is in the system.NUMCPUS
: Tells how many CPUs are present in the system.
It is possible to add the following to the config file:
. $SHELLPACK_INCLUDE/include-sizes.sh
get_numa_details
This will give access to more information about the system topology, such as:
NUMLLCS
: Number of Last Level Caches present in the system.NUMNODES
: Number of NUMA nodes.
Benchmark configurations can then be refined, by taking advantage of the knowledge of the platform characteristics.
For an example check
config-workload-stream-omp-llcs, where this is done: STREAM_THREADS=$NUMLLCS
. Or
config-scheduler-schbench,
which has: SCHBENCH_THREADS=$(((NUMCPUS/NUMNODES)-1))
The entry point to running benchmarks is run-mmtests.sh
. If run with -h
or --help
the available options are shown:
run-mmtests [-mnpb] [-c config-file] test-name
Options:
-m|--run-monitors Run with monitors enabled as specified by the configuration
-n|--no-monitor Only execute the benchmark, do not monitor it
-p|--performance Set the performance cpufreq governor before starting
-c|--config Configuration file to read (default: ./config)
-b|--build-only Only build the benchmark, do not execute it
If no config file is specified, the one in ./config
is used.
After a run, the benchmark results as well as any data that can be useful for
a report will be available in the work/log
directory. (more specifically,
in work/log/TEST_RUN/iter-0
, for a run called TEST_RUN
).
Note that often a configuration will run more than just one single benchmark
(this depends on the value of the MMTESTS
option in the config itself),
resulting in some subdirectories being present in the results directory.
Configuring the system for running a benchmark may include doing some changes
to the system itself that can only be done with root
provileges.
For instance, the config-workload-thpchallenge-fio-defrag config does:
echo always > /sys/kernel/mm/transparent_hugepage/defrag
If starting run-mmtests.sh
as a "regular user" doing something like that
will fail. benchmarks should still complete (most likely with some warnings)
but results will likely not be the ones expected.
In fact, MMTests is intended to be run as root
. For most of the changes that
it applies to the system, the framework is careful to (try to) undo them. It
is however fair to say that MMTests is best used on machines that can be
redeployed and reset to a clean known state both before and after running a
benchmark.
A full list of available monitors is in monitors/
.
The following options, to be defined in the config file, can be used to control monitoring:
RUN_MONITOR
:yes
orno
switch for deciding whether monitoring should happen or not, during the execution of the benchmarks. If set tono
, even if monitors are defined, they will be ignored (but seeMONITOR_ALWAYS
below). It can be overridden by the--run-monitors
and--no-monitor
command line parameters. I.e.,--run-monitors
means we will always run monitors, even if we haveRUN_MONITOR=no
in the config file (and vice versa, for--no-monitor
andRUN_MONITOR=yes
).MONITORS_ALWAYS
: Basically, another override. In fact, monitors defined here will be started even if we haveRUN_MONITOR=no
and/or--no-monitor
.MONITORS_GZIP
: A list of monitors to be used during the benchmarks. Their output will be saved in compressed (with gzip) log files.MONITORS_WITH_LATENCY
: A list of monitors to be used during the benchmarks with their output augmented with some additional timestamping.MONITOR_UPDATE_FREQUENCY
: How frequently, in seconds, the various defined monitors should produce and log a sample.MONITOR_FTRACE_OPTIONS
,MONITOR_FTRACE_EVENTS
: respectively, options to set and tracing events to enable forftrace
, if the "ftrace" monitor is enabled.MONITOR_PERF_EVENTS
: list ofperf
events to stat or record, when any of the "perf-foo" monitor is enabled (see below).
The files in monitors/
all follow the same naming scheme, which is
watch-foo.[sh|pl]
. For instance, we have monitors/watch-mpstat.sh
,
monitors/watch-proc-interrupts.sh
monitors/watch-proc-vmstat.sh
. For
monitoring the output of mpstat
and the content of /proc/interrupts
and
/proc/vmstat
during the execution of a benchmark, include this option in
to the config file: MONITORS_GZIP="proc-vmstat mpstat proc-interrupts"
Similarly, to monitor the output of vmstat
and iostat
, and also add
some timestamps to the output, define this option:
MONITORS_WITH_LATENCY="vmstat iostat"
.
In order to record the output of, for instance, the sched_migrate_task
tracepoint, make sure to have ftrace
in the list of monitors defined in
MONITORS_GZIP
and then add
MONITOR_FTRACE_EVENTS="sched/sched_migrate_task"
.
(See also configs/config-monitor-vm-stalls
for a more advanced example.)
For using perf
"as a monitor", a list of events should be defined, e.g.
MONITOR_PERF_EVENTS=cpu-migrations,context-switches
or
MONITOR_PERF_EVENTS=node-load-misses,node-store-misses
. Also, the monitor
should be defined either adding perf-time-stat
to the list of
MONITORS_GZIP
, or adding perf-event-stat
to the MONITORS_TRACER
option.
For reporting, there is a basic compare-kernels.sh
script.
Despite the name, it can compare an arbitrary number of benchmarking runs. The name has historical reasons, from the time when the only use case was comparing kernel versions, but nowadays anything can be compared --machines, userspace packages, benchmark versions, tuning parameters etc.
It is optionally possible to specify a different baseline and comparison points, while by default the results are organised by the time the test was executed.
NAME
compare-kernels.sh - Compare results between benchmarking runs
SYNOPSIS
compare-kernels.sh [options]
Options:
--baseline <testname> Baseline test name, default is time ordered
--compare "<test> <test>" Comparison test names, space separated
--exclude "<test> <test>" Exclude test names
--auto-detect Attempt to automatically highlight significant differences
--sort-version Assume kernel versions for test names and attempt to sort
--format html Generate a HTML format of the report
--output-dir Output directory for HTML report
It must be run from within an MMTests results directory. So, even if the
benchmarks have been run on a different machine, it is enought to capture
work/log
and run compare-kernels.sh
from there.
In the table(s) produced, it is usually the most interesting to look at the average values, computed over the individual results of multiple repetitions of the benchmarks. Note that some benchmarks use the harmonic mean (Hmean) and some use the arithmetic mean (Amean), depending of the nature of the results.
compare-kernel.sh
can generate an HTML report, with both tables and graphs.
For doing that, both the format and the output directory needs to be
specified. The HTML page will then come directly out of the standard output
of the tool. Therefore, invoking it like this is recommended:
$ cd work/log
$ mkdir /tmp/report
$ ../../compare-kernels.sh --format html --output-dir /tmp/report > /tmp/report/index.html
An example of the HTML reporting is available
here.
This comes from two simple runs of the default config
(i.e., of the
STREAM benchmark) when the system was idle (TEST_RUN
) and busy with
something else (TEST_RUN_BUSY
).
It is possible to obtain a report using a different tool. It is the script
that compare-kernels.sh
calls internally and it is located at
bin/compare-mmtests.pl
.
The output is the same table(s) produced by compare-kernels.sh
.
A possible invocation could look like this:
./bin/compare-mmtests.pl --directory work/log --benchmark stream --names TEST_RUN,TEST_RUN_BUSY
TEST_RUN TEST_RUN_BUSY
MB/sec copy 19059.86 ( 0.00%) 15234.88 ( -20.07%)
MB/sec scale 14078.10 ( 0.00%) 11258.38 ( -20.03%)
MB/sec add 14740.32 ( 0.00%) 11749.84 ( -20.29%)
MB/sec triad 14504.22 ( 0.00%) 11317.26 ( -21.97%)
If the benchmark does multiple operations —like STREAM above that checks the
memory throughput of four different operations— there will be one result for
each. In these cases, compare-mmtests.pl
can be used to produce an overall
comparison between the benchmarks.
This is done by taking the geometric mean (Gmean) of the results. The geometric mean is chosen because it has the nice property that the mean of ratios is equal to the ratios of the means, so we do not get different results depending on the order of the operations.
Looking at the Gmean offers a concise and hence rather useful overview of the overall performance, especially when complex benchmarks are used.
For instance:
./bin/compare-mmtests.pl --directory work/log/ --benchmark stream --names TEST_RUN,TEST_RUN_BUSY --print-ratio
TEST_RUN TEST_RUN_BUSY
Ratio copy 1.00 (0.00%) (NaNs) 0.80 (-20.07%) (NaNs)
Ratio scale 1.00 (0.00%) (NaNs) 0.80 (-20.03%) (NaNs)
Ratio add 1.00 (0.00%) (NaNs) 0.80 (-20.29%) (NaNs)
Ratio triad 1.00 (0.00%) (NaNs) 0.78 (-21.97%) (NaNs)
Gmean Higher 1.00 0.79
Of course, the Gmean for the benchmark chosen as the baseline will always
be 1.00
. Additionally, the Higher
or Lower
label tells us whether
it is the higher or lower values that represent better performance.
In the example above, TEST_RUN_BUSY
reaches only the 79% of TEST_RUN
performance, which means that it is 21% slower.
Further info about reporting:
The install and test scripts are automatically generated from "shellpacks".
A shellpack is a pair of benchmark and install scripts that are stored in
shellpacks/
.
Actual shellpacks are automatically generated from template files stored
in shellpack_src/src/
. Some have a build suffix indicating that it is only
building a supporting tool like a library a benchmark requires. Do not
modify the generated test-scripts in shellpacks/
directory as they will
simply be overwritten.
For instance, /shellpacks/shellpack-bench-pgbench
—which will be
automatically generated from shellpack_src/src/pgbench/pgbench-bench
—
contains all the individual test steps.
Each test is driven by bin/run-single-test.sh
script which reads
the relevant drivers/driver-<testname>.sh
script (e.g.,
drivers/driver-pgbench.sh
).
MMTests needs to download the various benchmarks from their official location, i.e., from the Internet. That might be problematic because it can (should!) be considered not trusted, or even just because the official repository may have been updated to a newer version which maybe is not yet compatible with the current release of MMTests. And if this happens, the run will likely fail.
Other potential problems are that the download may fail due to temporary networking issues, that it consumes bandwidth and that it adds delays and makes testing longer.
It is therefore possible to create a local mirror. The location of such mirror
can be configured in WEBROOT
in shellpacks/common-config.sh
.
For example, kernbench
tries to download $WEBROOT/kernbench/linux-3.0.tar.gz
.
If this is not available, it is downloaded from the internet.
This can add delays in testing and consumes bandwidth so is worth configuring.
Patches should be sent to Mel Gorman <mgorman@techsingularity.net>. While the project is hosted on github, notifications get lost so pull requests there may be missed for quite a long time.
(Pseudo-)Random links to when MMTests got mentioned around in the Internet:
- MMTests being used to benchmark patches to the task wake-up path inside the Linux scheduler, on LKML here and here.
- MMTests used to reproduce a bug in the accounting code inside the Linux scheduler, on LKML.
- MMTests used to benchmark some early version of the Core Scheduling patches, highlighting their impact on both baremetal and virtualization workloads, on LKML (check the replies for seeing all the benchmark results).
- Additionally to the above examples, a lot more reports of MMTests being used for Linux kernel development can be found just by searching for 'MMTests' in an LKML archive.
- Giovanni Gherdovich explaining running MMTests and reading the reporting on LKML.
Talks and presentation about or related to MMTests:
- Scheduler benchmarking with MMTests is a report of a talk about MMTests given at 2020 OSPM conference (slides).
- FOSDEM 2020 talk about MMTests, focusing on using it for running benchmarks inside virtual machines Automated Performance Testing for Virtualization with MMTests.
- Mel Gorman's talk at SUSE Labs Conference 2018, Marvin: Automated assistant for development and CI. It's about Marvin, but mentions MMTests as well.
- Jan Kara's talk at Open Source Summit 2017, Detecting Performance Regressions in the Linux Kernel.
- Jan Kara's talk at SUSE Labs Conference 2017, Performance Team's Grid.
- Davidlohr's talk at LinuxCon NA 2015 Performance Monitoring in the Linux Kernel.
Some historic references: