Interbench - The Linux Interactivity Benchmark


	Introduction

This benchmark application is designed to benchmark interactivity in Linux. See
the file readme.interactivity for a brief definition. 

It is designed to measure the effect of changes in Linux kernel design or system
configuration changes such as cpu, I/O scheduler and filesystem changes and
options. With careful benchmarking, different hardware can be compared.


	What does it do?

It is designed to emulate the cpu scheduling behaviour of interactive tasks and
measure their scheduling latency and jitter. It does this with the tasks on
their own and then in the presence of various background loads, both with
configurable nice levels and the benchmarked tasks can be real time.


	How does it work?

First it benchmarks how best to reproduce a fixed percentage of cpu usage on the
machine currently being used for the benchmark. It saves this to a file and then
uses this for all subsequent runs to keep the emulation of cpu usage constant.

It runs a real time high priority timing thread that wakes up the thread or
threads of the simulated interactive tasks and then measures the latency in the
time taken to schedule. As there is no accurate timer driven scheduling in linux
the timing thread sleeps as accurately as linux kernel supports, and latency is
considered as the time from this sleep till the simulated task gets scheduled.

Each benchmarked simulation runs as a separate process with its own threads,
and the background load (if any) also runs as a separate process.


	What interactive tasks are simulated and how?

X:
X is simulated as a thread that uses a variable amount of cpu ranging from 0 to
100%. This simulates an idle gui where a window is grabbed and then dragged
across the screen.

Audio:
Audio is simulated as a thread that tries to run at 50ms intervals that then
requires 5% cpu. This behaviour ignores any caching that would normally be done
by well designed audio applications, but has been seen as the interval used to
write to audio cards by a popular linux audio player. It also ignores any of the
effects of different audio drivers and audio cards. Audio is also benchmarked
running SCHED_FIFO if the real time benchmarking option is used.

Video:
Video is simulated as a thread that tries to receive cpu 60 times per second
and uses 40% cpu. This would be quite a demanding video playback at 60fps. Like
the audio simulator it ignores caching, drivers and video cards. As per audio,
video is benchmarked with the real time option.

Gaming:
The cpu usage behind gaming is not at all interactive, yet games clearly are
intended for interactive usage. This load simply uses as much cpu as it can
get. It does not return deadlines met as there are no deadlines with an
unlocked frame rate in a game. This does not accurately emulate a 3d game
which is gpu bound (limited purely by the graphics card), only a cpu bound
one.

Custom:
This load will allow you to specify your own combination of cpu percentage and
intervals if you have a specific workload you are interested in and know the
cpu usage and frame rate of it on the hardware you are testing.


	What loads are simulated?

None:
Otherwise idle system.

Video:
The video simulation thread is also used as a background load.

X:
The X simulation thread is used as a load.

Burn:
A configurable number of threads fully cpu bound (4 by default).

Write:
A streaming write to disk repeatedly of a file the size of physical ram.

Read:
Repeatedly reading a file from disk the size of physical ram (to avoid any
caching effects).

Compile:
Simulating a heavy 'make -j4' compilation by running Burn, Write and Read
concurrently.

Memload:
Simulating heavy memory and swap pressure by repeatedly accessing 110% of
available ram and moving it around and freeing it. You need to have some
swap enabled due to the nature of this load, and if it detects no swap this
load is disabled.

Hack:
This repeatedly runs the benchmarking program "hackbench" as 'hackbench 50'.
This is suggested as a real time load only but because of how extreme this
load is it is not unusual for an out-of-memory kill to occur which will
invalidate any data you get. For this reason it is disabled by default.

Custom:
The custom simulation is used as a load.


	What is measured and what does it mean?

1. The average scheduling latency (time to requesting cpu till actually getting
it) of deadlines met during the test period. 
2. The scheduling jitter is represented by calculating the standard deviation
of the latency
3. The maximum latency seen during the test period
4. Percentage of desired cpu
5. Percentage of deadlines met.

This data is output to console and saved to a file which is stamped with the
kernel name and date. See sample.log.

	Sample:
--- Benchmarking simulated cpu of X in the presence of simulated ---
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.495 +/- 0.495         45		 100	         96
Video	   11.7 +/- 11.7        1815		89.6	       62.7
Burn	   27.9 +/- 28.1        3335		78.5	         44
Write	   4.02 +/- 4.03         372		  97	       78.7
Read	   1.09 +/- 1.09         158		99.7	         88
Compile	   28.8 +/- 28.8        3351		78.2	       43.7
Memload	   2.81 +/- 2.81         187		98.7	         85

What can be seen here is that never during this test run were all the so called
deadlines met by the X simulator, although all the desired cpu was achieved
under no load. In X terms this means that every bit of window movement was
drawn while moving the window, but some were delayed and there was enough time
to catch up before the next deadline. In the 'Burn' column we can see that only
44% of the deadlines were met, and only 78.5% of the desired cpu was achieved.
This means that some deadlines were so late (%deadlines met was low) that some
redraws were dropped entirely to catch up. In X terms this would translate into
jerky movement, in audio it would be a skip, and in video it would be a dropped
frame. Note that despite the massive maximum latency of >3seconds, the average
latency is still less than 30ms. This is because redraws are dropped in order
to catch up usually by these sorts of applications.


	What is relevant in the data?

The results pessimise quite a lot what happens in real world terms because they
ignore the reality of buffering, but this allows us to pick up subtle 
differences more readily. In terms of what would be noticed by the end user,
dropping deadlines would make noticable clicks in audio, subtle visible frame
time delays in video, and loss of "smooth" movement in X. Dropping desired cpu
would be much more noticeable with audio skips, missed video frames or jerks
in window movement under X. The magnitude of these would be best represented by
the maximum latency. When the deadlines are actually met, the average latency
represents how "smooth" it would look. Average humans' limit of perception for
jitter is in the order of 7ms. Trained audio observers might notice much less.


	How to use it?

In response to critisicm of difficulty in setting up my previous benchmark, 
contest, I've made this as simple as possible.

	Short version:
make
./interbench -L XXX

where XXX is the number CPUs in your system.

Please read the long version before submitting results!

	Longer version:
Build with 'make'. It is a single executable once built so if you desire to
install it simply copy the interbench binary wherever you like.

To get good reproducible data from it you should boot into runlevel one so
that nothing else is running on the machine. All power saving (cpu throttling,
cpu frequency modifications) must be disabled on the first run to get an
accurate measurement for cpu usage. You may enable them later if you are
benchmarking their effect on interactivity on that machine. Root is almost
mandatory for this benchmark, or real time privileges at the very least. You
need free disk space in the directory it is being run in the order of 2* your
physical ram for the disk loads. A default run in v0.21 takes about 15
minutes to complete, longer if your disk is slow.

As the benchmark bases the work it does on the speed of the hardware the
results from different hardware can not be directly compared. However changes
of kernels, filesystem and options can be compared. To do a comparison of
different cpus and keep the workload constant, using the -l option and
passing the value of "loops_per_ms" from the first hardware tested will keep
the number of cpu cycles fairly constant allowing some comparison. Future
versions may add the option of setting the amount of disk throughput etc.


Command line options supported:
interbench [-l <int>] [-L <int>] [-t <int] [-B <int>] [-N <int>]
        [-b] [-c] [-r] [-C <int> -I <int>] [-m <comment>]
        [-w <load type>] [-x <load type>] [-W <bench>] [-X <bench>]
        [-h]

 -l     Use <int> loops per sec (default: use saved benchmark)
 -L     Use cpu load of <int> with burn load (default: 4)
 -t     Seconds to run each benchmark (default: 30)
 -B     Nice the benchmarked thread to <int> (default: 0)
 -N     Nice the load thread to <int> (default: 0)
 -b     Benchmark loops_per_ms even if it is already known
 -c     Output to console only (default: use console and logfile)
 -r     Perform real time scheduling benchmarks (default: non-rt)
 -C     Use <int> percentage cpu as a custom load (default: no custom load)
 -I     Use <int> microsecond intervals for custom load (needs -C as well)
 -m     Add <comment> to the log file as a separate line
 -w     Add <load type> to the list of loads to be tested against
 -x     Exclude <load type> from the list of loads to be tested against
 -W     Add <bench> to the list of benchmarks to be tested
 -X     Exclude <bench> from the list of benchmarks to be tested
 -h     Show help

There is one hidden option which is not supported by default, -u
which emulates a uniprocessor when run on an smp machine. The support for cpu
affinity is not built in by default because there are multiple versions of
the sched_setaffinity call in glibc that not only accept different variable
types but across architectures take different numbers of arguments. For x86
support you can change the '#if 0' in interbench.c to '#if 1' to enable the
affinity support to be built in. The function on x86_64 for those very keen
does not have the sizeof argument.


Thanks:
For help from Zwane Mwaikambo, Bert Hubert, Seth Arnold, Rik Van Riel,
Nicholas Miell, John Levon, Miguel Freitas and Peter Williams.
Aggelos Economopoulos for contest code, Bob Matthews for irman (mem_load)
code, Rusty Russell for hackbench code and Julien Valroff for manpage.

Sat Oct 31 15:17:12 2009
Con Kolivas < kernel at kolivas dot org >