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 ] [-L ] [-t ] [-N ] [-b] [-c] [-r] [-C -I ] [-m ] [-w ] [-x ] [-W ] [-X ] [-h] -l Use loops per sec (default: use saved benchmark) -L Use cpu load of with burn load (default: 4) -t Seconds to run each benchmark (default: 30) -B Nice the benchmarked thread to (default: 0) -N Nice the load thread to (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 percentage cpu as a custom load (default: no custom load) -I Use microsecond intervals for custom load (needs -C as well) -m Add to the log file as a separate line -w Add to the list of loads to be tested against -x Exclude from the list of loads to be tested against -W Add to the list of benchmarks to be tested -X Exclude 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 >