1 Interbench - The Linux Interactivity Benchmark
6 This benchmark application is designed to benchmark interactivity in Linux. See
7 the file readme.interactivity for a brief definition.
9 It is designed to measure the effect of changes in Linux kernel design or system
10 configuration changes such as cpu, I/O scheduler and filesystem changes and
11 options. With careful benchmarking, different hardware can be compared.
16 It is designed to emulate the cpu scheduling behaviour of interactive tasks and
17 measure their scheduling latency and jitter. It does this with the tasks on
18 their own and then in the presence of various background loads, both with
19 configurable nice levels and the benchmarked tasks can be real time.
24 First it benchmarks how best to reproduce a fixed percentage of cpu usage on the
25 machine currently being used for the benchmark. It saves this to a file and then
26 uses this for all subsequent runs to keep the emulation of cpu usage constant.
28 It runs a real time high priority timing thread that wakes up the thread or
29 threads of the simulated interactive tasks and then measures the latency in the
30 time taken to schedule. As there is no accurate timer driven scheduling in linux
31 the timing thread sleeps as accurately as linux kernel supports, and latency is
32 considered as the time from this sleep till the simulated task gets scheduled.
34 Each benchmarked simulation runs as a separate process with its own threads,
35 and the background load (if any) also runs as a separate process.
38 What interactive tasks are simulated and how?
41 X is simulated as a thread that uses a variable amount of cpu ranging from 0 to
42 100%. This simulates an idle gui where a window is grabbed and then dragged
46 Audio is simulated as a thread that tries to run at 50ms intervals that then
47 requires 5% cpu. This behaviour ignores any caching that would normally be done
48 by well designed audio applications, but has been seen as the interval used to
49 write to audio cards by a popular linux audio player. It also ignores any of the
50 effects of different audio drivers and audio cards. Audio is also benchmarked
51 running SCHED_FIFO if the real time benchmarking option is used.
54 Video is simulated as a thread that tries to receive cpu 60 times per second
55 and uses 40% cpu. This would be quite a demanding video playback at 60fps. Like
56 the audio simulator it ignores caching, drivers and video cards. As per audio,
57 video is benchmarked with the real time option.
60 The cpu usage behind gaming is not at all interactive, yet games clearly are
61 intended for interactive usage. This load simply uses as much cpu as it can
62 get. It does not return deadlines met as there are no deadlines with an
63 unlocked frame rate in a game. This does not accurately emulate a 3d game
64 which is gpu bound (limited purely by the graphics card), only a cpu bound
68 This load will allow you to specify your own combination of cpu percentage and
69 intervals if you have a specific workload you are interested in and know the
70 cpu usage and frame rate of it on the hardware you are testing.
73 What loads are simulated?
76 Otherwise idle system.
79 The video simulation thread is also used as a background load.
82 The X simulation thread is used as a load.
85 A configurable number of threads fully cpu bound (4 by default).
88 A streaming write to disk repeatedly of a file the size of physical ram.
91 Repeatedly reading a file from disk the size of physical ram (to avoid any
95 Simulating a heavy 'make -j4' compilation by running Burn, Write and Read
99 Simulating heavy memory and swap pressure by repeatedly accessing 110% of
100 available ram and moving it around and freeing it. You need to have some
101 swap enabled due to the nature of this load, and if it detects no swap this
105 This repeatedly runs the benchmarking program "hackbench" as 'hackbench 50'.
106 This is suggested as a real time load only but because of how extreme this
107 load is it is not unusual for an out-of-memory kill to occur which will
108 invalidate any data you get. For this reason it is disabled by default.
111 The custom simulation is used as a load.
114 What is measured and what does it mean?
116 1. The average scheduling latency (time to requesting cpu till actually getting
117 it) of deadlines met during the test period.
118 2. The scheduling jitter is represented by calculating the standard deviation
120 3. The maximum latency seen during the test period
121 4. Percentage of desired cpu
122 5. Percentage of deadlines met.
124 This data is output to console and saved to a file which is stamped with the
125 kernel name and date. See sample.log.
128 --- Benchmarking simulated cpu of X in the presence of simulated ---
129 Load Latency +/- SD (ms) Max Latency % Desired CPU % Deadlines Met
130 None 0.495 +/- 0.495 45 100 96
131 Video 11.7 +/- 11.7 1815 89.6 62.7
132 Burn 27.9 +/- 28.1 3335 78.5 44
133 Write 4.02 +/- 4.03 372 97 78.7
134 Read 1.09 +/- 1.09 158 99.7 88
135 Compile 28.8 +/- 28.8 3351 78.2 43.7
136 Memload 2.81 +/- 2.81 187 98.7 85
138 What can be seen here is that never during this test run were all the so called
139 deadlines met by the X simulator, although all the desired cpu was achieved
140 under no load. In X terms this means that every bit of window movement was
141 drawn while moving the window, but some were delayed and there was enough time
142 to catch up before the next deadline. In the 'Burn' column we can see that only
143 44% of the deadlines were met, and only 78.5% of the desired cpu was achieved.
144 This means that some deadlines were so late (%deadlines met was low) that some
145 redraws were dropped entirely to catch up. In X terms this would translate into
146 jerky movement, in audio it would be a skip, and in video it would be a dropped
147 frame. Note that despite the massive maximum latency of >3seconds, the average
148 latency is still less than 30ms. This is because redraws are dropped in order
149 to catch up usually by these sorts of applications.
152 What is relevant in the data?
154 The results pessimise quite a lot what happens in real world terms because they
155 ignore the reality of buffering, but this allows us to pick up subtle
156 differences more readily. In terms of what would be noticed by the end user,
157 dropping deadlines would make noticable clicks in audio, subtle visible frame
158 time delays in video, and loss of "smooth" movement in X. Dropping desired cpu
159 would be much more noticeable with audio skips, missed video frames or jerks
160 in window movement under X. The magnitude of these would be best represented by
161 the maximum latency. When the deadlines are actually met, the average latency
162 represents how "smooth" it would look. Average humans' limit of perception for
163 jitter is in the order of 7ms. Trained audio observers might notice much less.
168 In response to critisicm of difficulty in setting up my previous benchmark,
169 contest, I've made this as simple as possible.
175 Please read the long version before submitting results!
178 Build with 'make'. It is a single executable once built so if you desire to
179 install it simply copy the interbench binary wherever you like.
181 To get good reproducible data from it you should boot into runlevel one so
182 that nothing else is running on the machine. All power saving (cpu throttling,
183 cpu frequency modifications) must be disabled on the first run to get an
184 accurate measurement for cpu usage. You may enable them later if you are
185 benchmarking their effect on interactivity on that machine. Root is almost
186 mandatory for this benchmark, or real time privileges at the very least. You
187 need free disk space in the directory it is being run in the order of 2* your
188 physical ram for the disk loads. A default run in v0.21 takes about 15
189 minutes to complete, longer if your disk is slow.
191 As the benchmark bases the work it does on the speed of the hardware the
192 results from different hardware can not be directly compared. However changes
193 of kernels, filesystem and options can be compared. To do a comparison of
194 different cpus and keep the workload constant, using the -l option and
195 passing the value of "loops_per_ms" from the first hardware tested will keep
196 the number of cpu cycles fairly constant allowing some comparison. Future
197 versions may add the option of setting the amount of disk throughput etc.
200 Command line options supported:
201 interbench [-l <int>] [-L <int>] [-t <int] [-B <int>] [-N <int>]
202 [-b] [-c] [-r] [-C <int> -I <int>] [-m <comment>]
203 [-w <load type>] [-x <load type>] [-W <bench>] [-X <bench>]
206 -l Use <int> loops per sec (default: use saved benchmark)
207 -L Use cpu load of <int> with burn load (default: 4)
208 -t Seconds to run each benchmark (default: 30)
209 -B Nice the benchmarked thread to <int> (default: 0)
210 -N Nice the load thread to <int> (default: 0)
211 -b Benchmark loops_per_ms even if it is already known
212 -c Output to console only (default: use console and logfile)
213 -r Perform real time scheduling benchmarks (default: non-rt)
214 -C Use <int> percentage cpu as a custom load (default: no custom load)
215 -I Use <int> microsecond intervals for custom load (needs -C as well)
216 -m Add <comment> to the log file as a separate line
217 -w Add <load type> to the list of loads to be tested against
218 -x Exclude <load type> from the list of loads to be tested against
219 -W Add <bench> to the list of benchmarks to be tested
220 -X Exclude <bench> from the list of benchmarks to be tested
223 There is one hidden option which is not supported by default, -u
224 which emulates a uniprocessor when run on an smp machine. The support for cpu
225 affinity is not built in by default because there are multiple versions of
226 the sched_setaffinity call in glibc that not only accept different variable
227 types but across architectures take different numbers of arguments. For x86
228 support you can change the '#if 0' in interbench.c to '#if 1' to enable the
229 affinity support to be built in. The function on x86_64 for those very keen
230 does not have the sizeof argument.
234 For help from Zwane Mwaikambo, Bert Hubert, Seth Arnold, Rik Van Riel,
235 Nicholas Miell, John Levon, Miguel Freitas and Peter Williams.
236 Aggelos Economopoulos for contest code, Bob Matthews for irman (mem_load)
237 code, Rusty Russell for hackbench code and Julien Valroff for manpage.
239 Sat Mar 4 12:11:34 2006
240 Con Kolivas < kernel at kolivas dot org >