Check out the new USENIX Web site.

Home About USENIX Events Membership Publications Students
USENIX '05 Paper    [USENIX '05 Technical Program]

Auto-pilot: A Platform for System Software Benchmarking

Auto-pilot: A Platform for System Software Benchmarking

Charles P. Wright, Nikolai Joukov, Devaki Kulkarni, Yevgeniy Miretskiy, and Erez Zadok
Stony Brook University


When developing software, it is essential to evaluate its performance and stability, making benchmarking an essential and significant part of the software development cycle. Benchmarking is also used to show that a system is useful or provide insight into how systems behave. However, benchmarking is a tedious task that few enjoy, but every programmer or systems researcher must do. Developers need an easy-to-use system for collecting and analyzing benchmark results. We introduce Auto-pilot, a tool for producing accurate and informative benchmark results. Auto-pilot provides an infrastructure for running tests, sample test scripts, and analysis tools. Auto-pilot is not just another metric or benchmark: it is a system for automating the repetitive tasks of running, measuring, and analyzing the results of arbitrary programs. Auto-pilot can run a given test until results stabilize, automatically highlight outlying results, and automatically detect memory leaks. We have used Auto-pilot for over three years on eighteen distinct projects and have found it to be an invaluable tool that saved us significant effort.

1  Introduction

Benchmarking contributes evidence to the value of work, lends insight into the behavior of systems, and provides a mechanism for stress-testing software. However, benchmarking can be an arduous task. Benchmarking takes a lot of time, and the initial iteration of benchmarks often exposes bugs or inefficient code. After changes are made to the code, the benchmarks need to be repeated, time and time again. Once the code is bug free and stable, there are sometimes unexpected results. The output then needs to be examined to determine a cause. If no cause for the suspicious results is found, then the benchmarks need to be repeated without changing any parameters to verify the unexpected results. After the results are verified, one variable needs to be changed at a time to narrow down the source of the erroneous data. This cycle of benchmarking and analysis is often repeated many times.
We have identified two primary considerations when collecting and analyzing benchmark results:
Accuracy The numbers that benchmarks produce need to be mathematically correct, but an even more difficult requirement to satisfy is that they need to be reproducible, stable, and fair. They should be reproducible so that you can re-run the test and get similar results. This way if you need to make slight modifications, it is possible to go back and compare results. To help achieve this, Auto-pilot records pertinent system information (e.g., the OS version, hard disk model, and partitions) so that you can reproduce the test environment as closely as possible. Tests also need to be reproducible so that others can verify your results. For example, each test should be run under similar circumstances (e.g., a cold or warm cache). If one test fails, then it may impact all future tests, so a series of tests should be stopped so that erroneous results are not included.
Presentation The accuracy of results is inconsequential if the results can not be understood and correctly interpreted. Each benchmark usually results in hundreds or thousands of numbers. For example, each execution of a program results in elapsed, user, and system time. Additionally, the CPU utilization and wait (I/O) time are reported. If there is a four-threaded benchmark that is run twenty times, then 400 values are produced. There are also usually many configurations of a benchmark. Each variable that is introduced multiplicatively increases this value. In this example, if we ran the test for 1, 2, 4, 8, 16 and 32 threads, there would be a total of 6,300 values. For every additional quantity that is measured, another 1,260 numbers are generated. Without assistance it is rather tedious to wade through this sea of numbers.
Auto-pilot is an infrastructure that produces accurate and informative results. Auto-pilot is not a metric or a specific benchmark, but rather a framework to run benchmarks. Auto-pilot includes a language for describing a series of benchmarks using a simple syntax that includes basic loops and conditionals. The Auto-pilot distribution includes a set of scripts for several compile benchmarks and also for running multiple concurrent Postmark processes [12]. Auto-pilot also provides useful analysis tools that handle multi-process benchmarks that can also account for background processes or kernel threads. Auto-pilot can automatically stop the benchmarks after reliable results are obtained, highlight outlying values, and detect memory leaks. Results are presented in a tabular format that can easily be imported into spreadsheets. Auto-pilot also includes a bar and line graph script that generates graphs from the tabular results using Gnuplot.
One of the most frustrating tasks when benchmarking is deciding how many times to run a test. The ideal situation is to run the benchmark precisely as many times as needed to obtain stable results, but not any more than that (to save time for both the benchmarker and to free the testbed for more tests). Auto-pilot can automatically determine when test results have met an arbitrary stability condition (e.g., the half-width of the confidence interval is within 5% of the mean) and stop the tests at that point.
Benchmarking is often used to debug software. Software may leave some state behind that affects future results. For example, a kernel module may have small memory leaks that eventually exhaust system resources. Auto-pilot uses linear regression to determine if there are memory leaks (which reduce the total amount of free memory) , or a performance degradation after each iteration of the benchmark.
Even software that works correctly changes the state of the system. For example, a working set of files is loaded into OS caches. Often, researchers unmount and remount the file system to provide cold cache results, but rebooting has measurable advantages over simply remounting the file system. Unfortunately, rebooting usually adds time-consuming manual intervention to the benchmarking process. To ensure a consistent system state, Auto-pilot provides checkpointing support. Checkpointing and rebooting is fully automated: all Auto-pilot state is written to a file, the machine restarts, and after system initialization completes Auto-pilot resumes the benchmarks from where it left off.
Systems have a long life cycle and benchmark results may need to be analyzed months or years after they were taken (e.g., after a paper is accepted, reviewers often ask questions about the performance evaluation). Complete results must be saved for future analysis in a meaningful way. Auto-pilot carefully stores relevant system information and all program output.
The rest of this paper is organized as follows. We describe previous work in Section 2. In Section 3 we describe the design and implementation of Auto-pilot. We conclude in Section 4.

2  Background

There are many metrics and benchmarks available to test systems. In this section we describe a few notable systems and how they differ from Auto-pilot.
Several tools seek to measure precisely (down to microseconds, nanoseconds, or CPU cycles) the amount of time it takes to perform a single operation. Two examples of these tools are lmbench [14] and hbench:OS [5]. These tools include both a set of tests, and also analysis tools for these specific tests. For example, lmbench measures latency and bandwidth for memory, IPC, file system operations, disk I/O, cached I/O, TCP, UDP, and RPC. Several O/S primitives are also measured by lmbench, including system call entry, context switching, signal handling, process creation, and program execution (fork+exec). These precise measurements are useful to debug parts of a system, but they do not measure the interactions between system calls that exist in more realistic workloads. The lmbench suite contains not only the infrastructure to run and report these benchmarks, but defines the tests as well. In contrast, for Auto-pilot we focus on running relatively large scale tests (whole programs), like Postmark, a compile, or micro-benchmarks that perform some operation many times. For example, we have used Auto-pilot to benchmark specific file system operations like stat and readdir by running find over a tree of files.
Brown and Seltzer developed hbench:OS, which is a modified version of lmbench that improves timing and statistical methodology, adds more parameters to tests, and improves individual benchmarks. In lmbench many tests are run in a loop and a final result is calculated based on all the runs, but some tests are only run once. Because some architectures have coarse-grained timing infrastructures, running the test may produce inaccurate results (e.g., 0 microseconds to perform an operation like a TCP connect). To remove this deficiency, hbench:OS uses a self-scaling loop that runs the test for at least one second, which is several orders of magnitude more than even the worst timing mechanisms. For tests that can be run only once, hbench:OS uses CPU cycle counters. Different methods are used to report results in lmbench and hbench:OS. For some lmbench tests a mean is reported, for others a minimum. In hbench:OS each individual measurement is recorded so that data analysis is separated from reporting. In hbench:OS, n% trimmed means are used for all results. The lowest and highest n% results are discarded, and the remaining (100 - 2n%) results are used to compute an arithmetic mean. Whether trimmed means are a better method for results analysis is disputed by the lmbench authors, but because hbench:OS stores raw results, different types of analysis are still possible. Overall, hbench:OS and Auto-pilot are different for the same reasons lmbench and Auto-pilot are different: we focus on large general purpose benchmarks, whereas these two systems focus on small micro-benchmarks. However, in Auto-pilot we have made some similar decisions to hbench:OS for reporting and analysis. In Auto-pilot, we record all results and test output. Auto-pilot is also flexible with support for arbitrary metrics (e.g., elapsed time, I/O operations, or packets sent). Our analysis tools can then operate on the raw output to produce reports and graphs. Like hbench:OS, Auto-pilot can also automatically scale benchmarks to the testbed (e.g., run a benchmark for one hour or until the confidence interval's half-width meets a threshold).
Profilers and other tools measure how long specific sections of code are executed (e.g., functions, blocks, lines or instructions) [3,9]. Profilers can be useful because they often tell you how to make your program faster, but the profiling itself often changes test conditions and adds overhead, making it unsuitable for comparing the performance of two systems. The Perl Benchmark::Timer module [10] times sections of Perl code within a script. Before the code section, a start function is called; and after the section an end function is called. Benchmark::Timer can run a code section for a fixed number of times, or alternatively until the width of a confidence interval meets some specified value. Auto-pilot can use similar methodology to determine how many iterations of a test should be run.
Other benchmarks like SDET [7] and AIM7 [1] are system-level benchmarks. Both SDET and AIM7 run a pre-configured workload with increasing levels of concurrency. The metric for each benchmark is the peak throughput. These systems focus on developing a metric, and measuring that specific metric, not running arbitrary benchmarks.
The closest system to Auto-pilot is Software Testing Automation Framework (STAF) [11] developed by IBM. STAF is an environment to run specified test cases on a peer-to-peer network of machines. Rather than measuring the performance of a given test, STAF aims to validate that the test case behaved as expected. STAF runs as a service on a network of machines. Each machine has a configuration file that describes what services the other machines may request it to perform (e.g., execute a specific program). STAF also provides GUI monitoring tools for tests. The major disadvantage with STAF is that it requires complex setup, does not focus on performance measurement, and is a heavy-weight solution for running multiple benchmarks on a single machine
The Open Source Development Lab (OSDL), provides a framework called the Scalable Test Platform (STP) that allows developers to test software patches on systems with 1-8 processors [13]. STP allows developers to submit a patch for testing, and then automatically deploys the patch on a system, executes the test, and posts the results on a Web page. STP makes it relatively simple to add benchmarks to the framework, but the benchmarks themselves need to be changed to operate within the STP environment. Auto-pilot is different from STP in two important ways. First, STP is designed for many users to share a pool of machines, whereas Auto-pilot is designed to repeatedly run a single researcher's set of tests on a specific machine. Second, STP provides no analysis tools; it simply runs tests and logs the output, whereas Auto-pilot measures processes and provides tools to analyze results.

3  Design

To run a benchmark, you must write a configuration file that describes which tests to run and how many times. The configuration file does not describe the benchmark itself, but rather points at another executable. This executable is usually a small wrapper shell script that provides arguments to a program like Postmark or a compile benchmark. The wrapper script is also responsible for measurement. We provide sample configuration files and shell scripts for benchmarking file systems. These can be run directly for common file systems, or easily adapted for other types of tests. Given a configuration file and shell scripts, the next step is to run the configuration file with Auto-pilot. Auto-pilot parses the configuration file and runs the tests, producing two types of logs. The first type is simply the output from the programs. This can be used to verify that benchmarks executed correctly and to investigate any anomalies. The second log file is a more structured results file that contains a snapshot of the system and the measurements that were collected. The results file is then passed through our analysis program, Getstats, to create a tabular report. Optionally, the tabular report can be used to generate a bar or line graph.
In Section 3.1 we describe auto-pilot, the Perl script that runs the benchmarks and logs the results. In Section 3.2 we describe the sample shell scripts specific to the software being benchmarked. In Section 3.3 we describe Getstats, which produces summaries and statistical reports of the Auto-pilot output. In Section 3.4 we describe our plotting scripts. In Section 3.5 we describe and evaluate checkpointing and resuming benchmarks across reboots. In Section 3.6 we describe using hooks within the benchmarking scripts to benchmark NFS.

3.1  auto-pilot

The core of the Auto-pilot system is a Perl script that parses and executes the benchmark scripts. Each line of a benchmark script contains a command (blank lines and comments are ignored). The command interface has been implemented to resemble the structure of a typical programming language. This makes Auto-pilot easy and intuitive to enhance when writing new benchmarks. Next, we describe the thirteen primary commands.
TEST   begins a benchmark. The test directive takes between two and four arguments. The first argument is the name of the test, which is also used to build the name of the output file. The second argument is the minimum number of times to run the test. The third and fourth arguments are optional. The fourth argument specifies a program that determines if the test should continue. If the program's exit status is zero, then the benchmark stops otherwise it continues. The third argument is how many iterations of the benchmark to execute before rerunning the program specified in the fourth argument. This argument is useful because it may be more efficient to execute several iterations of the benchmark between runs of the program rather than running a potentially computationally intensive program after a single iteration. Using a program to determine if tests should continue allows the benchmark to execute until there are stable results, thereby saving time on the testbed, and leaving more time for analysis. Getstats, described in Section 3.3, has support for arbitrary predicates to determine stability. The test directive contains SETUP, EXEC, and CLEANUP directives (described below). Other control directives are also allowed (e.g., IF directives, variable declarations, and loops).
THREADS   tells Auto-pilot how many concurrent benchmark processes should be run at once. This can be used to test the scalability of systems. Our analysis tools, described in Section 3.3, aggregate the results from these threads.
EXEC   executes a test. If the THREADS directive is used, then threads processes are simultaneously executed. The results for each thread are logged to separate files. Auto-pilot sets the environment variable APTHREAD to 1 in the first thread, 2 in the second thread, and N in the n-th thread. This allows each thread to perform slightly different tasks (e.g., performing tests in different directories).
SETUP   executes a setup script for a test. Setup scripts are not multi-threaded, and are used for initialization that is common to all threads (e.g., to mount a test file system). We also support a PRESETUP directive that is run only once. One possible use of PRESETUP is to format a file system only once, and remount it before each test using a SETUP directive.
CLEANUP   scripts are used to undo what is done in a setup script. CLEANUP scripts can be used to ensure a cold cache for the next test. To run a script after all the iterations of the test are completed, a POSTCLEANUP directive can be used for final cleanup.
VAR   sets an Auto-pilot variable. Simple variable substitution is performed before executing each line: %VAR% is replaced with the value of VAR. These variables are not exposed to external processes.
ENV   sets an environment variable and the corresponding Auto-pilot variable. This can be used to communicate with benchmark scripts without the need for many command-line arguments. Auto-pilot also replaces $VAR$ with the contents of the environment variable VAR. This is similar to the export command within the Bourne shell.
IF   is a basic conditional that supports equality, greater-than, and less-than (and their inverses). If the condition evaluates to true, then all statements until ELSE or FI are executed. If the condition is false, then the optional statements between ELSE and FI are executed. ELSE-IF blocks are also supported.
WHILE   repeatedly executes a block of code while a certain condition holds true. The condition syntax WHILE uses is the same as IF.
FOREACH   assigns multiple values to one variable in turn. This is useful because often a single test needs to be repeated with several different configurations.
FOR   is similar to FOREACH, but instead of specifying each value explicitly, a start, end, increment, and factor are specified. For example, FOR THREADCOUNT=1 TO 32 FACTOR 2 would execute the loop with a THREADCOUNT of 1, 2, 4, 8, 16, and 32.
FASTFAIL   causes Auto-pilot to abort if one of the benchmarks may not be successful. If the benchmarks continue, then they may destroy important state that could lend insight into the cause of the failure. An optional fastfail script is also defined, which can be used to send email to the person responsible for the benchmarks. We have found it very useful to email pagers, so that testbeds do not remain idle after a failed benchmark.
CHECKPOINT   writes all Auto-pilot internal state to a file. If Auto-pilot is invoked with the checkpoint file as an argument, then it resumes execution from where it left off. The return value of CHECKPOINT is similar to the Unix fork system call. After restoring a checkpoint, the value of the Auto-pilot variable RESTORE is 1, but after writing the checkpoint, the value of RESTORE is 0. In Section 3.5 we present and evaluate an in-depth example of checkpointing across reboots.
Configuration example   Figure 1 shows the Postmark configuration file included with Auto-pilot.
2  FOREACH FS ext2 ext3 reiserfs
5      TEST %FS%:%THREADS% 10 1 getstats \
  --predicate '("$delta" < 0.05 * $mean) \
  || ($count > 30)'
6        SETUP %FS%
7        EXEC
8        CLEANUP %FS%
9      DONE
10   DONE
Figure 1: A sample Auto-pilot configuration file
Line 1 includes, a configuration file that performs actions that are common to all tests. also turns off all excess services (e.g., Cron) to prevent them from interfering with a benchmark run. The file also includes, which the user can create to set variables including TESTROOT, which defines the directory where the test will take place. Line 2 begins a FOREACH loop. This loop is executed a total of three times. The first time the value of FS is set to "ext2," the second time it is set to "ext3," and on the third it is "reiserfs." Line 3 begins a nested loop that will set the THREADCOUNT variable to 1, 2, 4, 8, 16, and 32. Line 4 sets the number of threads to use to THREADCOUNT. Lines 5-9 define the test. The test is named %FS%:%THREADS%. For example, the first test is named ext2:1. The test is executed 10 times, and then after each test the Getstats program (described in Section 3.3), is used to determine whether the confidence interval has an acceptable half-width for elapsed, user, and system time. The test is stopped after 30 runs, to prevent a test from running forever. Line 6 calls with an argument of ext2. This is our file system setup script, which formats and mounts a file system (formatting can be disabled with environment variables). Line 7 executes This shell script creates a Postmark configuration, and then executes Postmark through our measurement facility (described in Section 3.2). Line 8 unmounts the Ext2 file system, so that the next run takes place with a cold cache. Lines 9-11 close their corresponding loops. Line 12 turns services back on, and optionally sends an email to the user indicating that the benchmarks are complete.

3.2  Benchmark Scripts

We provide a set of file system benchmarking scripts for Postmark [12] and compiling various packages. We included an example script for compiling Am-Utils [15], GCC [8], and OpenSSL [17]. Most other packages can be compiled by setting environment variables or with minimal changes to the existing scripts.
Many applications can be benchmarked without any scripts, but file systems require complex setup and cleanup. Our scripts also serve as an example for benchmarking applications in other domains. We wrote scripts to test some of our complex file systems. These systems required extensive testing and Auto-pilot allowed us to methodically test and debug them.
The flow of the scripts is organized as two components. The first component mounts and unmounts the file system. The second performs the Postmark or the compile benchmark. Though we distribute scripts for benchmarking compilations and Postmark on Ext2, Ext3, and Reiserfs, it is easy to add other benchmarks or platforms. To test additional file systems, the Auto-pilot scripts have hooks for mount, unmount, mkfs, tunefs, and more. With these hooks, new file systems or new file system features can be used. We provide example hooks for enabling HTrees on Ext2/3 [16] and to benchmark on top of stackable file systems [21]. We describe how we used the hooks to concurrently benchmark an NFS client and server in Section 3.6.
The sample Auto-pilot scripts we distribute demonstrate the following principles:
  • Separating the test and the setup. Most benchmarks involve performing several workloads on several configurations. Each configuration has a setup and a corresponding cleanup script. Each workload has a script common to all configurations.
  • Using variables for all values that may change. We additionally have support for operating system and host-specific options.
  • Unmounting the file system on which the benchmark takes place between runs, even if the previous run failed.
Auto-pilot includes, which calls to shut down Cron, Sendmail, Anacron, LPD, Inetd, and other services on the testbed. If no user is logged in via SSH, then SSH is also turned off. We create /etc/nologin to prevent users from logging in while the test is being run. Swap space is optionally disabled to prevent it from affecting results. All these actions are taken to avoid unexpected user or system activity from distorting the results.
The file system setup script loads in machine-specific settings (e.g., which device and directory to use). Next, the script logs some vital statistics about the machine (OS version, CPU information, memory usage, hard disk configuration, and partition layout). The script then unmounts any previously-mounted file systems, formats the device, and mounts the new file system (e.g., Ext2, Ext3, or Reiserfs). This ensures a clean cache and consistent disk layout. Because vital machine information is recorded, the test conditions can be reproduced in the future.
Next, the benchmark is executed. The compile script unpacks, configures, and compiles the package specified. The compile commands are run through a function named ap_measure. This function produces a block of information that encapsulates the results for this test. By default, the block includes the iteration of the test, which thread was running, the elapsed time, the system CPU time, the user CPU time, and the command's exit status. When analyzing the results, this block logically becomes a row in a spreadsheet. Additional fields can be added through the use of measurement hooks. Measurement hooks are executed before and after the test. Each hook can produce values to be included in the block. We provide sample hooks that measure the number of I/O requests for a given partition; the amount of free memory; the amount of CPU time used by background processes; network utilization; and more. Programs such as CFS [4] or Cryptoloop [18] use a separate thread for some processing. Even if the instrumented process itself does not have additional threads, the kernel may use asynchronous helper threads to perform certain tasks (e.g., bdflush to flush dirty buffers or kjournald to manage journalized block devices). This hook allows measurement of daemons that expend effort on behalf of the measured process. The measurement hook is also called one final time so that it can produce its own blocks. This feature is used for the CPU time difference measurement hook to enumerate all processes that used additional CPU time. Our I/O measurement script also adds a block with the number of I/O operations that occurred on each partition.
The Postmark script is more complicated because it can run several Postmark processes in parallel. When running a multithreaded test, Auto-pilot sets two environment variables: APTHREAD and APIPCKEY. APTHREAD is the thread number for this test. APTHREAD can be used, for example, so that different processes use different directories. APIPCKEY is used for synchronization. First, the Postmark script creates a directory and then sets up a Postmark configuration file. Because Auto-pilot starts each test sequentially, one test could start before another and then the timing results would be inaccurate. To solve this, Auto-pilot creates a System V semaphore with the number of threads as its value. Before executing Postmark, each test script decrements the semaphore. We include a small C utility that calls semctl to decrement the semaphore. All of the test processes are suspended until the semaphore reaches zero, at which point they all begin to execute concurrently.
After all the benchmark processes complete, Auto-pilot executes the cleanup script, which unmounts the file system. After all tests are executed successfully, SSH is restarted and /etc/nologin is removed. This is not meant to restore the machine to its previous state, but rather to allow remote access, so that the benchmarker can retrieve results, or begin another series of benchmarks. Auto-pilot does not keep track of which services it stopped. To restore the services the machine was running before the benchmarks, it can simply be rebooted.
ext2:1.res: High z-score of 2.21853230276562 for elapsed in epoch 3.
Elapsed     10  6.055  6.063  5.991  6.120  5.855  6.180 1.491 1.067
System      10  2.758  2.760  2.709  2.807  2.640  2.880 2.499 1.788
User        10  1.675  1.680  1.615  1.735  1.510  1.820 5.044 3.609
Wait        10  1.622  1.636  1.567  1.677  1.465  1.718 4.759 3.404
CPU%        10 73.221 73.079 72.572 73.871 72.007 74.981 1.240 0.887

Elapsed     15 77.861 76.865 74.156 81.567 64.308 88.209 8.594 4.759 1185.869
System      15  4.272  4.290  4.217  4.327  4.100  4.410 2.334 1.293   54.895
User        15  1.825  1.820  1.773  1.877  1.670  1.990 5.132 2.842    8.935
Wait        15 71.765 70.775 68.064 75.466 58.158 82.179 9.312 5.157 4324.025
CPU%        15  7.885  7.923  7.499  8.272  6.836  9.563 8.850 4.901  -89.231
Figure 2: The Getstats tabular report format showing the results for Postmark with a single thread on Ext2 and Ext3.

3.3  Getstats

Getstats is a Perl script which processes the results log file to generate useful tabular reports. In addition to Auto-pilot results files, Getstats can process Comma Separated Value (CSV) files or the output of GNU time. Getstats is flexible in that it does not hardcode the types of information it expects in these files; it simply reads and displays the data. The parsers themselves are also modular. Getstats searches the Perl library path for valid Getstats parsers and loads them. To write a new parser, two functions must be defined: a detection function and a parsing function that reads the input file into a two-dimensional array. We originally used this functionality to add support for GNU time files, but have also used it for some of our own custom formats. We discuss mostly timing information, but we have also analyzed network utilization, I/O operations, memory, and other quantities with Getstats.
Getstats has a basic library of functions to transform the data. Examples of transformations include adding a column derived from previous values, selecting rows based on a condition, raising warnings, grouping data based on the value of a column, or aggregating data from multiple rows to produce a single summary statistic. If Getstats detects that it is being run on results files with time data, it performs some default transformations composed of the basic library transformations. These include raising warnings if a command failed, aggregating multiple threads into a single value, computing a wait time (the time the measured process was not running) and a CPU-utilization column, raising warnings if any test had a high z-score for one if its values, computing overheads, and finally generating a tabular report.
Figure 2 shows a tabular report generated by Getstats compares Ext2 with Ext3, when run on the Postmark configuration discussed in Section 3.1. The first line of the report is a high z-score warning for the third iteration of ext2 with one thread. If there are tests with very large z-scores, then there likely were problems with the benchmark. A few other high z-score warnings were issued, but are not shown to conserve space. Of note is that Ext2 ran only ten times, but Ext3 needed to run 15 times to get an acceptable half-width percentage for the measured quantities (the predicates are not run for Wait and CPU% which are computed quantities). This is not unexpected because journaling adds complexity to the file system's I/O pattern, resulting in greater variability [6]. The tabular format itself is useful as a mechanism to present detailed results or to import data into spreadsheets, but graphs present a better overall picture of the results.
Output Modes   Getstats provides several useful options to analyze and view different output modes. The simplest mode outputs the raw uninterpreted values, which is useful when there are problems with the benchmark (e.g., one run had anomalous results). The tabular report also has information useful for plotting. The LOW and HIGH columns are suitable for creating error bars with Gnuplot. Getstats supports several methods of creating error bars: Student-t confidence intervals, the minimum and maximum value, or the standard deviation. By default Getstats reports the count, mean, median, minimum, maximum, and Student-t confidence interval error bar values (shown as low and high), and the standard deviation and half-width of the confidence interval as a percent of the mean. The standard deviation is a measure of how much variance there is in the tests. The half-width of the confidence interval describes how far the true value may be from the mean with a given degree of confidence (by default 95%). If multiple results files are specified, then the first file is used as a baseline to compute overheads for the subsequent files (this is shown as "O/H"). This report gives a high-level yet concise overview of a test.
Elapsed 10    38.649 38.193 37.950 39.348 37.673 40.379 2.528 1.808 
System  10    1.663  1.675  1.603  1.723  1.540  1.770  5.071 3.628 
Elapsed 10    38.751 38.699 38.580 38.921 38.465 39.307 0.614 0.439 0.262  
System  10    1.796  1.790  1.677  1.915  1.580  2.080  9.255 6.620 7.998  
Comparing remount (Sample 1) to chill (Sample 2).
Elapsed: 95%CI for remount - chill = (-0.567, 0.769)
Null Hyp. Alt. Hyp. P-value Result     
u1 <= u2  u1 >  u2  0.377   ACCEPT H_0 
u1 >= u2  u1 <  u2  0.623   ACCEPT H_0 
u1 == u2  u1 != u2  0.754   ACCEPT H_0 

System: 95%CI for remount - chill = (0.009, 0.257)
Null Hyp. Alt. Hyp. P-value Result     
u1 <= u2  u1 >  u2  0.018   REJECT H_0 
u1 >= u2  u1 <  u2  0.982   ACCEPT H_0 
u1 == u2  u1 != u2  0.037   REJECT H_0 
Figure 3: The Getstats t-test output. User time, Wait time, and CPU usage are omitted for brevity.
Statistical Tests   After changing your software you would like to know if your changes actually had a measurable effect on performance (or some other measured quantity). In some cases it is sufficient to compare the means, and if they are "close", then you may assume that they are the same or that your change did not noticeably affect performance. In other circumstances, however, a more rigorous approach should be used. For example, if two tests are very close, it can be difficult to determine if there is indeed a difference, or how large that difference really is. To compare two samples, Getstats can compute the confidence interval for the difference between the means, and can also run a two-sample t-test. The confidence interval quickly tells you if there is a difference, and how much it is. The confidence interval is simply a range of numbers. If that range includes zero, then the samples are not significantly different. Getstats also can run a two-sample t-test to determine the relationship between two results files. A statistical test has a null hypothesis, which is assumed to be true. An example of a null hypothesis is u1 = u2. The result of a t-test is a P-value, which is the probability that you would observe the data if the null hypothesis is true. If the P-value is large (close to 1), then your data is consistent with the null hypothesis. If the P-value is small (closer to 0), then your data is not consistent with the null hypothesis. If the P-value is smaller than a predetermined significance level (e.g., .05), then you reject the original assumption (i.e., the null hypothesis).
Figure 3 shows the output of a Getstats t-test for two samples: CHILL and REMOUNT. We ran a recursive grep benchmark over the GCC 3.4.3 source tree. For CHILL we ran a program that we wrote called chill that is designed to ensure cold-cache results. Our version of chill was inspired by a similar program in SunOS 4 [20]. Chill allocates and dirties as much memory as possible, thereby forcing the kernel to evict unused objects. We hypothesized that chill would provides more stable results than simply unmounting the test file system, because it causes all caches to be purged, not just those related to the test file system. The REMOUNT configuration unmounts the file system and then mounts it again instead of running chill.
From the tabular report, we can see that elapsed time differs by only 0.262%, and system time differs by only 7.998%. After the tabular report, each measured value in the samples is compared (e.g, elapsed time from the first sample is compared to elapsed time from the second sample).
If we examine the confidence interval for the elapsed time, we can see that it includes zero (the beginning of the range is less than zero, but the end of the range is greater than zero), which means that CHILL and REMOUNT are not distinguishable. The next four lines show the results of the t-test for elapsed time. Getstats runs the t-test with three distinct assumptions: u1 < = u2, u1 > =u2, and u1 = u2. For simplicity, we chose to display all of the tests rather than forcing the user to specify the tests of interest. It is up to the benchmarker to determine which of these assumptions makes sense. In all cases, the P-value was higher than the significance level of (0.05), which means that we can not reject the assumptions with 95% confidence. Therefore, we conclude that CHILL and REMOUNT are indistinguishable for elapsed time.
For system time, the confidence interval does not contain zero, which means the two tests have a significant difference. This is supported by the P-value, which shows us that there is only a 3.7% chance of observing this data if the two samples were in fact equal. There is also only a 1.8% chance that CHILL uses less system time than REMOUNT. Because 3.7% and 1.8% are less than the significance level, we can reject the null hypothesis that u1 = u2 and that u1 u2 where u1 represents REMOUNT and u2 represents CHILL. Because our assumption that REMOUNT used less system time than CHILL is false, we know that REMOUNT did use more system time because it needed to recreate objects that the kernel evicted.
Detecting Anomalies   Often, one or two bad runs are not noticeable from summary statistics alone. Getstats provides two mechanisms for finding anomalous results:
  • Automatically highlighting outlying values
  • Performing linear regression on the values
To automatically highlight outlying values, Getstats uses the z-score of each point. The z-score is the difference between the value and the mean, divided by the standard deviation. If the z-score is greater than a configurable value (by default 2), a warning is printed before the summary statistics. The benchmarker can then look into the results further to analyze the cause of the outliers.
Getstats can also compute a least-squares linear regression to fit the elapsed time, system time, user time, and free memory. Linear regression computes the slope and intercept of the line that best fits a set of points. If the results are stable, then the slope of these should be zero. The intercept is just the magnitude of the values, so it is unimportant. If the magnitude of the slope is not close to zero, then a warning is issued. If the free memory slope is negative, then it indicates that on each iteration the software is leaking memory. If the elapsed or CPU times have positive slopes, it similarly indicates that some resource is becoming more scarce, and causing a gradual performance degradation. Depending on the system being analyzed, different thresholds are appropriate, so the default value can be overridden by the benchmarker. To ensure that the amount of free memory reported is accurate, we run our version of chill. Chill allocates as much memory as possible, then dirties every page. Chill is terminated by the Linux out-of-memory manager (or when malloc can no longer get additional memory). At this point the kernel has evicted objects and pages to make room for chill, and the amount of memory consumed should be at a minimum.
Predicate Evaluation   Getstats evaluates predicates for use with auto-pilot's TEST directive. Getstats runs the predicate over each column in the results file, and returns zero if the predicate is true. The predicates are flexible. Getstats replaces several statistical variables (e.g., mean, median, half-width of the confidence interval, standard deviation, number of tests, and the slope of the linear regression) with their actual values and then passes the predicate to Perl's eval function. This combination of substitution and eval allows arbitrarily complex predicates using simple math or Perl functions. For example, the statement "0.05 * $mean < $delta" returns true if the half-width of the Student-t confidence interval is less than 5% of the mean.
Combining Results   Benchmarks often have too many numbers for a user to interpret, so Auto-pilot automatically combines results in two specific cases:
  • Benchmarks that consist of multiple programs (e.g., compilations)
  • Multi-threaded benchmarks
Some benchmarks consist of multiple commands. For example, to compile GCC, three steps are required: tar extracts the distribution, ./configure detects information about the environment, and finally make builds the package. Getstats unifies the results of each command into one result per benchmark execution. Alternatively, the user can use a select transformation to analyze one specific command in isolation.
When analyzing multi-threaded data, Getstats aggregates all of the threads in one test together to create a single set of results for the test. The elapsed time that is reported is the longest running thread and the system and user times are the sum of the system and user times for each thread. This allows a benchmark that is usually single-threaded to be run and analyzed in a multi-threaded manner.

3.4  Plotting Tools

One of the most useful ways to present benchmark data is through the use of graphs. In our experience, it is desirable to have automatically-generated graphs in Encapsulated Postscript (EPS) format, so the graphs are more suitable for publication. Our requirement for automatic graph generation essentially dictates that we use a command-line tool. We have found Gnuplot to be the most flexible plotting package, but its interface is cumbersome, often requiring dozens of lines of code and a specific data format to create a simple plot. Auto-pilot uses a wrapper script, Graphit, that generates both line and bar charts using Gnuplot. Although Gnuplot is our preference, everyone has their own favorite plotting tool. Each plotting wrapper script is specific to the plotting tool used, but Getstats can produce CSV output which can easily be imported into other programs (e.g., Excel).
Figure 4: The results for Postmark on Reiserfs, 1-32 threads
A sample line graph can be seen in Figure 4. This graph shows the results from Postmark running on top of Reiserfs with 1-32 threads. The elapsed time decreases as threads are added, because the kernel can better schedule I/O operations. However, the user and system CPU time remained relatively constant because the same amount of work is being done.
This graph was generated with the following command line:

graphit -mode=line \
  -components=Elapsed,User,System \
  -graphfile=reiserfs-pm.eps \
  -k 'top right' -f 17 \
  -xlabel Threads -ylabel "Time (seconds)"  \
  Reiserfs 'reiserfs:*.res'

The mode argument specifies a line graph (the other supported mode is a bar graph). The -components parameter informs Graphit which quantities are of interest. The -graphfile parameter specifies an output file. The next three parameters are optional: -k specifies the legend location (Gnuplot refers to the legend as a key); -f increases the font size to 17 points; and xlabel and ylabel specify axis labels. The last two parameters define a series. In this case it is named Reiserfs, and the values are from the files named reiserfs:*.res. This command line replaces an 11 line Gnuplot script, with 566 characters (a savings of 70%). More importantly, Graphit creates a properly-formatted Gnuplot data file for the actual series.
Creating bar graphs in Gnuplot is not straightforward. Rather than defining series and values as is done in other plotting tools, you must create an artificial x-axis and locate each bar and its label along this axis. To properly space and locate these bars without automated tools is tedious. Graphit automatically constructs this artificial x-axis and spaces the bars appropriately. Using command-line arguments, the width of the bars, the gap between each bar, the gap between each group of bars, and several other parameters can be controlled easily. An example Graphit bar graph is shown in Figure 5. This graph shows the same Postmark results for Ext2, Ext3, and Reiserfs that are shown in Figures 2 and 4. Graphit automatically reads the results files, stacks user CPU time over system CPU time; determines bar widths, spacing, and error bars; formats a Gnuplot data file; generates a Gnuplot script; and finally runs Gnuplot. We ran the following command to generate the graph:

graphit -mode=bar \
  -ylabel `Time (seconds)' -f 17 \
  -components=Elapsed,User,System \
  -graphfile=pm.eps Ext2 ext2:1.res \
  Ext3 ext3:1.res Reiserfs reiserfs:1.res

This single command replaces a 15-line 764-character Gnuplot script, and, more importantly, generates a data file with appropriate spacing along the artificial axis. Graphit has several spacing options to control the width of each bar, the gap between bars (e.g., System and User could be next to each other), and the gap between sets of bars (e.g., Ext2 and Ext3). Figure 6 is another example Graphit bar graph using the same options but with more data sets, so the bars are thinner. We also passed "-rotate 45" to Graphit to rotate the X-axis labels.
Figure 5: Postmark on Ext2, Ext3, and Reiserfs.

3.5  Checkpointing Across Reboots

Operating system code behaves very differently when data is already cached (called a warm cache) and when data is not yet in the cache (called a cold cache). This concern is particularly acute for file systems, which depend on caches to avoid disk or network operations, which are orders of magnitude slower than in-memory operations. Creating a warm cache situation is relatively easy: to create n warm cache runs, you can run the test n+1 times in sequence and discard the first result, which is only used to warm the cache. However, testing a cold cache situation is more difficult. For a perfectly cold cache, all OS objects must be evicted from their caches between each test. For file systems research, researchers unmount the test file system between each test. This invalidates inodes, directory name lookup caches, and the page cache for the tested file system. Unfortunately, this approach has two key disadvantages. First, even though inodes (and other objects) are invalidated, they may not be deallocated. This can expose subtle bugs when inodes are not properly cleared before reuse. The second problem is that the kernel often keeps a pool of unused objects for faster allocation, so the second run may use less system time because it does not need to get raw pages for object caches. For these two reasons, we have additionally run our version of chill, described in Section 3.3, to deallocate these objects.
To get a truly cold cache, a reboot is required because all operating state is reset after a reboot. The disadvantage of a reboot is that it takes a long time and usually requires manual intervention. We have designed Auto-pilot so that it can serialize all of its state into a plain text file using the CHECKPOINT directive. The CHECKPOINT directive is similar to UNIX fork in that it sets a variable to "0" after writing a checkpoint, but the variable is set to "1" after resuming from where the checkpoint left off. After writing a checkpoint, an Auto-pilot script can reboot the machine. After the machine's initialization process completes (at the end of /etc/rc.d/rc.local), we check whether the checkpoint file exists. If the checkpoint file exists, is owned by root, and is not world-writable, then we start Auto-pilot and it resumes from where benchmark execution left off.
We wanted to quantify the differences between various methods of cooling the cache. We ran a recursive grep -q benchmark over the GCC 3.4.3 source tree. We chose grep because it is a simple read-oriented benchmark, yet has a significant user component. We used the six following configurations:
The benchmark is run repeatedly with a warm cache.
Chill is run between each iteration of the benchmark.
The test file system is remounted between each iteration of the benchmark.
The test file system is remounted between each iteration of the benchmark.
The machine is rebooted between each iteration of the benchmark.
The machine is rebooted and Chill is run between each iteration of the benchmark. We used this configuration because the initialization process after reboot may have caused some objects to be loaded into the cache, and chill may evict them.
All tests were run on a 1.7Ghz Pentium IV with Fedora Core 2 and a vanilla 2.4.23 kernel. The test partition was on a Western Digital 5,400 RPM IDE disk. Each test was run 10 times. The elapsed, system, and user time results are shown in Figure 6, and the error bars show the 95% confidence intervals for elapsed time.
Figure 6: grep -r with various cache-cooling methods.
In all cases user time differences were indistinguishable. This was expected, because grep's processing does not change with the different configurations. The Sequential configuration had the lowest elapsed and system time, because it does not need to perform any I/O to read objects from disk. For Chill and Remount, System time was not distinguishable, but Chill used 13% less system time, so remounting is a more effective way of cooling the cache. Remount+Chill is indistinguishable from remount. Reboot was 4.3% slower than remounting the file system, and system time was indistinguishable. Reboot is not distinguishable from Reboot+Chill. Therefore we conclude that rebooting the system is in fact the best method of cooling the cache.

3.6  Benchmarking Script Hooks for NFS

We have found that we continually enhance our Auto-pilots benchmarking scripts, yet each project often needs its own slightly different setup and cleanup mechanisms or needs to measure a new quantity. For example, our tracing file system needed to measure the size of the trace file that each test generated [2]. This made the scripts hard to maintain. Projects that were started earlier used older modified versions of the benchmarking scripts. As the scripts improved, these older projects would have progressively more out of date scripts. Recently, we have redesigned Auto-pilot's benchmarking scripts to be extensible and provide hooks in critical locations. Our goal was to allow developers to test new file systems and add new measurements to the benchmarking scripts, without having to modify the originals.
To do this, Auto-pilot automatically reads in all files in $APLIB/commonsettings.d ($APLIB is an environment variable that is set to the path where the scripts were installed, but additional paths can be added to it). These files define hooks for various events such as mounting a file system, unmounting a file system, beginning a measurement, and ending a measurement.
We prototyped a modified NFS client and server, and therefore did not use the standard NFS benchmark, SPEC-SFS [19], because it does not use the local client (it hand-crafts RPCs instead). To benchmark our modified client and server, we used Auto-pilot to run standard file system benchmarks, such as Postmark. We have included the hooks we used for this project as an example of what can be done with our hooks.
The simplest hook we have adds support for mounting an NFS file system. The hook requires two environment variables to be set (the server and the path to mount on the server). The hook also loads and unloads the NFS file system module (unless it is built into the kernel).
Because we were running a modified NFS server, we needed to coordinate with the server and measure the CPU time used by the NFS server (it does not make sense to measure elapsed time, because the server runs for precisely as long as the client). To coordinate between the client running the benchmarks and the server, we used SSH to run remote commands on the server. This follows the same model that SPEC SFS uses, but replaces RSH using rhosts with SSH using public-key based authentication. In the mount hook, we copy the module to the server and then restart it. In the unmount hook we unload the module. We also added a remote process measurement hook. Each time a client process is measured, we run a command on the NFS server to record the amount of CPU time used by nfsd. After the client process is terminated, we run another command on the server to report the difference in CPU time used.
The current scripts make it relatively easy to benchmark one client machine accessing one server, which is a relatively common benchmarking case. If you wanted to benchmark one server with multiple clients, then you would need to write new scripts in which the server runs Auto-pilot and remotely executes the benchmark on the clients. The server-side scripts would need to be written from scratch, but the client-side scripts could be very similar to the current Auto-pilot benchmarking scripts.

4  Conclusion

Auto-pilot provides a useful set of tools for accurately and informatively benchmarking software. Auto-pilot configurations are powerful scripts that describe a series of tests, including multi-threaded versions of traditional benchmarks. Auto-pilot includes sample scripts for various compile-based benchmarks and Postmark. Auto-pilot's flexible infrastructure allows many other tests and measurements to be added. For example, we include hooks for testing over NFS, measuring I/O operations and CPU time used by background processes. We have also used many different benchmarks aside from the ones included in the package (e.g., grep and other custom benchmarks).
We present results in an informative manner that can easily be used with Gnuplot. Getstats generates easy-to-read tabular reports, automatically displays outlying points, detects memory leaks, and runs statistical tests. In combination with Auto-pilot, Getstats can evaluate predicates to run tests until an arbitrarily complex condition is satisfied, thereby saving time on the testbed. Graphit processes results files and automatically creates Gnuplot scripts with properly formatted data files to create bar and line graphs.
Auto-pilot can checkpoint its state, reboot the machine, and resume running tests from where it left off. Rebooting provides a colder cache than simply remounting the file system, and requires no manual intervention on the part of the benchmarker. When benchmarks fail, Auto-pilot can automatically send email or text pages to the benchmarker to prevent the testbed from lying idle.
We have used Auto-pilot for over three years and for eighteen projects. Auto-pilot saved us many days of work in collecting performance results and reduced debugging cycles by exposing bugs more quickly.

Auto-pilot is released under the GPL and can be downloaded from Auto-pilot contains 5,799 lines of code. The Auto-pilot Perl script has 936 lines; Getstats has 1,647 lines; Graphit has 587 lines; the C utilities have 758 lines; the configuration files have 221 lines; and our shell scripts have 1,650 lines. All of our Perl scripts also have basic Perldoc formatted documentation that provides information on simple usage. We also include a full user manual that describes more detailed usage and has some brief tutorials in PDF and GNU info formats.

4.1  Future Work

We plan to add support for regular system snapshots, which can then be correlated with the output of the tests. This will allow a person conducting benchmarks to determine the context that the test was executed in. These snapshots will include kernel messages, memory usage, CPU usage, and other vital statistics.
Getstats currently uses a two-sample t-test to compare different results files. A t-test makes three assumptions:
  • The samples are independent
  • The samples are normally distributed
  • The samples have equal variance
Auto-pilot scripts attempt to ensure that each sample is independent of the other samples by purging cached data through remounting and rebooting. The last two assumptions must currently be verified by the benchmarker. If the sample size is sufficiently large (roughly greater than 30 samples), the t-test will still be appropriate. The third assumption must still be verified. We plan to automatically run an F-test on the variances. If the test concludes that the variances are not equal, then Getstats will use an approximate t-test for samples with unequal variances instead of the standard t-test.


Joseph Spadavecchia developed the first simple version of our benchmarking Perl script that has since morphed into the current Auto-pilot. Amit Purohit, Kiran-Kumar Muniswamy-Reddy, Michael Martino, Avishay Traeger, and others in our group have provided valuable feedback when using the system to benchmark their projects.
This work was partially made possible by NSF CAREER award EIA-0133589, NSF Trusted Computing Award CCR-0310493, and HP/Intel gift numbers 87128 and 88415.1.


AIM Technology. AIM Multiuser Benchmark - Suite VII Version 1.1., 2001.
A. Aranya, C. P. Wright, and E. Zadok. Tracefs: A File System to Trace Them All. In Proceedings of the Third USENIX Conference on File and Storage Technologies (FAST 2004), pages 129-143, San Francisco, CA, March/April 2004.
Bell Laboratories. prof, January 1979. Unix Programmer's Manual, Section 1.
M. Blaze. A cryptographic file system for Unix. In Proceedings of the first ACM Conference on Computer and Communications Security, 1993.
A. Brown and M. Seltzer. Operating System Benchmarking in the Wake of Lmbench: A Case Study of the Performance of NetBSD on the Intel x86 Architecture. In Proceedings of the 1997 ACM SIGMETRICS Conference on Measurement and Modeling of Computer Systems, pages 214-224. ACM Press, June 1997.
R. Bryant, R. Forester, and J. Hawkes. Filesystem Performance and Scalability in Linux 2.4.17. In Proceedings of the Annual USENIX Technical Conference, FREENIX Track, pages 259-274, Monterey, CA, June 2002.
S. Gaede. Perspectives on the SPEC SDET benchmark., January 1999.
The GCC team. GCC online documentation, 3.3.2 edition, August 2003.
S. L. Graham, P. B. Kessler, and M. K. McKusick. Gprof: A call graph execution profiler. In Proceedings of the 1982 SIGPLAN symposium on Compiler construction, pages 120-126, June 1982.
A. Ho and D. Coppit. Benchmark::Timer - Benchmarking with statistical confidence, December 2004. User Contributed Perl Documentation, Section 3.
IBM. Software testing automation framework STAF., 2001.
J. Katcher. PostMark: A New Filesystem Benchmark. Technical Report TR3022, Network Appliance, 1997.
Open Source Development Labs. Scalable test platform., 2004.
L. McVoy and C. Staelin. lmbench: Portable tools for performance analysis. In Proceedings of the USENIX 1996 Annual Technical Conference, pages 279-295, January 1996.
J. S. Pendry, N. Williams, and E. Zadok. Am-utils User Manual, 6.1b3 edition, July 2003.
D. Phillips. A directory index for EXT2. In Proceedings of the 5th Annual Linux Showcase & Conference, pages 173-182, November 2001.
The OpenSSL Project. Openssl: The open source toolkit for SSL/TLS., April 2003.
H. V. Riedel. The GNU/Linux CryptoAPI site., August 2003.
SPEC: Standard Performance Evaluation Corporation. SPEC SFS97_R1 V3.0., September 2001.
Sun Microsystems, Inc. Chill - remove useful pages from the virtual memory cache. SunOS 4 Reference Manual, Section 8.
E. Zadok and J. Nieh. FiST: A Language for Stackable File Systems. In Proceedings of the Annual USENIX Technical Conference, pages 55-70, San Diego, CA, June 2000.

File translated from TEX by TTH, version 3.61.
On 24 Feb 2005, 16:49.

This paper was originally published in the Proceedings of the 2005 USENIX Annual Technical Conference,
April 10–15, 2005, Anaheim, CA, USA

Last changed: 3 Mar. 2005 aw
USENIX '05 Technical Program
USENIX '05 Home