From Linux NFS
NFSometer is a performance measurement framework for running workloads and reporting results across NFS protocol versions, NFS options and Linux NFS client implementations.
The NFSometer project was started at NetApp as a way to automate performance testing of the Linux NFS client. Since then it has grown to include many other features, notably the generation of reports. It is our hope that by releasing this project under the GPLv2 license, the Linux NFS community will benefit from a better understanding of NFS performance characteristics and contribute improvements to NFSometer.
NFSometer is not designed to replace existing filesystem benchmarks. Instead, it is designed to automate the running of existing filesystem benchmarks, gather NFS specific statistics and generate reports that make benchmark output more understandable.
- Automate the time consuming process of running a variety of workloads looking for performance regressions between kernel versions, NFS protocol versions and NFS options.
- Generate reports that allows the user to understand the performance characteristics of workloads and easily compare performance characteristics of different configurations.
- Provide a simple way to define workloads. This allows users to evaluate NFS deployment scenarios with workloads representative of their unique needs.
- Wed Mar 6, 2013: NFSometer v1.5 released! This version includes several bugfixes, including one that caused a traceback when parsing results from older kernels.
- Fri Feb 22, 2013: NFSometer v1.4 released! This version has a redesign of reports, many bugfixes and several new features.
- Fri Dec 7, 2012: NFSometer v1.3 released! This version includes some major cleanup and bugfixes.
- Fri Nov 30, 2012: NFSometer v1.2 released! This version includes, new options (notably -o replaces -m), basic loadgen functionality, probe mountopts and detect tags, replace poorly supported admin-script and client-script functionality with tags (-t option), bugfixes, etc
- Thu Jun 21, 2012: NetApp is pleased to announce the GPLv2 release of NFSometer!
The current release of nfsometer is version 1.5:
nfsometer-1.5.tar.gz (md5sum: 4669fd1554ca51307c2f744e51a108a8)
- nfsometer-1.4.tar.gz (md5sum: ef6e3031b2c39c7eb439dd35197e6b76)
- nfsometer-1.3.tar.gz (md5sum: f6ee81e838753fe1d07e49fff653f5cf)
- nfsometer-1.2.tar.gz (md5sum: 86ca117b3265ce40408b2015e96482b5)
- nfsometer-1.1.tar.gz (md5sum: efbca65ec863b63cb852baac5ff409cf)
- nfsometer-1.0.tgz (md5sum: 9d8873517c0365dc04d28b585e0ba31b)
Main git repository:
'master' branch is current release, 'testing' branch is where all new development is happening.
- git clone git://git.linux-nfs.org/projects/dros/nfsometer.git
- NFSometer gitweb
Here is an example nfsometer report. The reports have notes explaining what was run.
example 1: run cthon workload with and without a loaded server.
Examples from nfsometer-1.4 manpage / "nfsometer examples":
Example 1: See what workloads are available $ nfsometer workloads This command lists available workloads and will tell you why workloads are unavailable (if any exist). Example 2: Compare cthon, averaged over 3 runs, across nfs protocol versions nfsometer -n 3 server:/export cthon This example uses the default for -o: "-o v3 -o v4 -o v4.1". To see the results, open results/index.html in a web browser. Example 3: Compare cthon, averaged over 3 runs, between v3 and v4.0 only nfsometer -n 3 -o v3 -o v4 server:/export cthon This example specifies v3 and v4 only. To see the results, open results/index.html in a web browser. Example 4: Compare two kernels running iozone workload, averaged over 2 runs, across all nfs protocol versions nfsometer can compare two (or more) kernel versions, but has no way of building, installing or booting new kernels. It's up to the user to install new kernels. In order for these kernels to be differentiated, 'uname -a' must be different. 1) boot into kernel #1 2) nfsometer -n 2 server:/export iozone 3) boot into kernel #2 4) nfsometer -n 2 server:/export iozone 5) open results/index.html in a web browser To see the results, open results/index.html in a web browser. Example 5: Using tags Tags (the -t option) can be used to mark nfsometer runs as occurring with some configuration not captured by mount options or detectable tags, such as different sysctl settings (client side), different server side options, or different network conditions. 1) set server value foo to 2.3 2) nfsometer -o v4 -o v4.1 -t foo=2.3 3) set server value foo to 10 4) nfsometer -o v4 -o v4.1 -t foo=10 What is passed to -t is entirely up to the user - it will not be interpreted or checked by nfsometer at all, so be careful! To see the results, open results/index.html in a web browser. Example 6: Always options The -o flag specifies distinct option sets to run, but sometimes there are options that should be present in each. Instead of writing each one out, you can use the -a option: nfsometer -o v3 -o v4 -a sec=krb5 server:/export iozone this is equivalent to: nfsometer -o v3,sec=krb5 -o v4,sec=krb5 server:/export iozone Example 7: Using the "custom" workload A main use case of nfsometer is the "custom" workload - it allows the user to specify the command that nfsometer is to run. NOTE: the command's cwd (current working directory) is the runroot created on the server. export NFSOMETER_CMD="echo foo > bar" export NFSOMETER_NAME="echo" export NFSOMETER_DESC="Writes 4 bytes to a file" nfsometer server:/export custom This will run 3 traces (v3, v4, v4.1) against server:/export of the command: echo foo > bar. Example 8: Using the loadgen mode Loadgen runs several instances of a workload without capturing traces. The idea is that you use several clients to generate load, then another client to measure performance of a loaded server. The "real" run of nfsometer (not loadgen) should mark the traces using the -t option. 1) On client A, run the cthon workload to get a baseline of a server without any load. nfsometer trace server:/export cthon 2) When that's done, start loadgen on client B: nfsometer -n 10 loadgen server:/export dd_100m_1k This runs 10 instances of dd_100m_1k workload on server:/export. It can take several minutes to start in an attempt to stagger all the workload instances. 3) once all instances are started, run the "real" nfsometer trace on client A. Use the -t option to mark the traces as having run under load conditions: nfsometer -t "10_dd" trace server:/export cthon 4) Explain how the tests were set up in the result notes. This should be run on client A (which has the traces: nfsometer notes 5) Now generate the reports: nfsometer report Example 8: Long running nfsometer trace The nfsometer.py script currently runs in the foreground. As such, it will be killed if the tty gets a hangup or the connection to the client is closed. For the time being, nfsometer should be run in a screen session, or run with nohup and the output redirected to a file. 1) screen -RD 2) nfsometer -n 2 server:/export iozone 3) close terminal window (or ^A^D) ... 4) reattach later with screen -RD 5) once nfsometer.py is done, results will be in results/index.html
Please send any bug reports, questions or comments to firstname.lastname@example.org
For contributions, see howto-contribute.txt.