• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1.. _faq:
2
3iperf3 FAQ
4==========
5
6What is the history of iperf3, and what is the difference between iperf2 and iperf3?
7  iperf2 was orphaned in the late 2000s at version 2.0.5, despite some
8  known bugs and issues. After spending some time trying to fix
9  iperf2's problems, ESnet decided by 2010 that a new, simpler tool
10  was needed, and began development of iperf3. The goal was make the
11  tool as simple as possible, so others could contribute to the code
12  base. For this reason, it was decided to make the tool single
13  threaded, and not worry about backwards compatibility with
14  iperf2. Many of the feature requests for iperf3 came from the
15  perfSONAR project (http://www.perfsonar.net).
16
17  Then in 2014, Bob (Robert) McMahon from Broadcom restarted
18  development of iperf2 (See
19  https://sourceforge.net/projects/iperf2/). He fixed many of the
20  problems with iperf2, and added a number of new features similar to
21  iperf3. iperf2.0.8, released in 2015, made iperf2 a useful tool. iperf2's
22  current development is focused is on using UDP for latency testing, as well
23  as broad platform support.
24
25  As of this writing (2017), both iperf2 and iperf3 are being actively
26  (although independently) developed.  We recommend being familiar with
27  both tools, and use whichever tool’s features best match your needs.
28
29  A feature comparison of iperf2, iperf3, and nuttcp is available at:
30  https://fasterdata.es.net/performance-testing/network-troubleshooting-tools/throughput-tool-comparision/
31
32iperf3 parallel stream performance is much less than iperf2. Why?
33  iperf3 is single threaded, and iperf2 is multi-threaded. We
34  recommend using iperf2 for parallel streams.
35  If you want to use multiple iperf3 streams use the method described `here <https://fasterdata.es.net/performance-testing/network-troubleshooting-tools/iperf/multi-stream-iperf3/>`_.
36
37I’m trying to use iperf3 on Windows, but having trouble. What should I do?
38  iperf3 is not officially supported on Windows, but iperf2 is. We
39  recommend you use iperf2.
40
41  Some people are using Cygwin to run iperf3 in Windows, but not all
42  options will work.  Some community-provided binaries of iperf3 for
43  Windows exist.
44
45How can I build a statically-linked executable of iperf3?
46  There are a number of reasons for building an iperf3 executable with
47  no dependencies on any shared libraries.  Unfortunately this isn't
48  quite a straight-forward process.
49
50  The steps below have nominally been tested on CentOS 7.4, but
51  can probably be adapted for use with other Linux distributions:
52
53  #.  If necessary, install the static C libraries; for CentOS this is
54      the ``glibc-static`` package.
55
56  #.  If OpenSSL is installed, be sure that its static libraries are
57      also installed, from the ``openssl-static`` package.
58
59  #.  Be sure that ``lksctp-*`` packages are not installed, because
60      as of this writing, there do not appear to be any static
61      libraries available for SCTP.
62
63  #.  Configure iperf3 thusly: ``configure "LDFLAGS=--static"
64      --disable-shared`` These options are necessary to disable the
65      generation of shared libraries and link the executable
66      statically.  For iperf-3.8 or later, configuring as ``configure
67      --enable-static-bin`` is another, shorter way to accomplish
68      this.  If SCTP is installed on the system it might also be
69      necessary to pass the ``--without-sctp`` flag at configure
70      time.
71
72  #.  Compile as normal.
73
74  It appears that for FreeBSD (tested on FreeBSD 11.1-RELEASE), only
75  the last two steps are needed to produce a static executable.
76
77How can I build on a system that doesn't support profiled executables?
78  This problem has been noted by users attempting to build iperf3 for
79  Android systems, as well as some recent versions of macOS.
80  There are several workarounds. In order from least
81  effort to most effort:
82
83  #. Beginning with iperf-3.8, profiled executables are actually not
84     built by default, so this question becomes somewhat moot.  Pass
85     the ``--enable-profiling`` flag to ``configure`` to build
86     profiled executables.
87
88  #. In iperf-3.6 and iperf-3.7, the ``--disable-profiling`` flag can be
89     passed to ``configure`` to disable the building of profiled
90     object files and the profiled executable.
91
92  #. At the time the linking of the iperf3 profiled executable fails,
93     the "normal" iperf3 executable is probably already created. So if
94     you are willing to accept the error exit from the make process
95     (and a little bit of wasted work on the build host), you might
96     not need to do anything.
97
98  #. After the configure step, there will be a definition in
99     ``src/Makefile`` that looks like this::
100
101       noinst_PROGRAMS = t_timer$(EXEEXT) t_units$(EXEEXT) t_uuid$(EXEEXT) \
102         iperf3_profile$(EXEEXT)
103
104     If you edit it to look like this, it will disable the build of the profiled iperf3::
105
106       noinst_PROGRAMS = t_timer$(EXEEXT) t_units$(EXEEXT) t_uuid$(EXEEXT)
107
108  #. Similar to item 2 above, but more permanent...if you edit
109     ``src/Makefile.am`` and change the line reading like this::
110
111       noinst_PROGRAMS         = t_timer t_units t_uuid iperf3_profile
112
113     To look like this::
114
115       noinst_PROGRAMS         = t_timer t_units t_uuid
116
117     And then run ``./bootstrap.sh``, that will regenerate the project
118     Makefiles to make the exclusion of the profiled iperf3 executable
119     permanant (within that source tree).
120
121I'm seeing quite a bit of unexpected UDP loss. Why?
122  First, confirm you are using iperf 3.1.5 or higher. There was an
123  issue with the default UDP send size that was fixed in
124  3.1.5. Second, try adding the flag ``-w2M`` to increase the socket
125  buffer sizes. That seems to make a big difference on some hosts.
126
127iperf3 UDP does not seem to work at bandwidths less than 100Kbps. Why?
128  You'll need to reduce the default packet length to get UDP rates of less that 100Kbps. Try ``-l100``.
129
130TCP throughput drops to (almost) zero during a test, what's going on?
131  A drop in throughput to almost zero, except maybe for the first
132  reported interval(s), may be related to problems in NIC TCP Offload,
133  which is used to offload TCP functionality to the NIC (see
134  https://en.wikipedia.org/wiki/TCP_offload_engine). The goal of TCP
135  Offload is to save main CPU performance, mainly in the areas of
136  segmentation and reassembly of large packets and checksum
137  computation.
138
139  When TCP packets are sent with the "Don't Fragment" flag set, which
140  is the recommended setting, segmentation is done by the TCP stack
141  based on the reported next hop MSS in the ICMP Fragmentation Needed
142  message. With TCP Offload, active segmentation is done by the NIC on
143  the sending side, which is known as TCP Segmentation offload (TSO)
144  or in Windows as Large Send Offload (LSO). It seems that there are
145  TSO/LSO implementations which for some reason ignore the reported
146  MSS and therefore don’t perform segmentation. In these cases, when
147  large packets are sent, e.g. the default iperf3 128KB (131,072
148  bytes), iperf3 will show that data was sent in the first interval,
149  but since the packets don’t get to the server, no ack is received
150  and therefore no data is sent in the following intervals. It may
151  happen that after certain timeout the main CPU will re-send the
152  packet by re-segmenting it, and in these cases data will get to the
153  server after a while. However, it seems that segmentation is not
154  automatically continued with the next packet, so the data transfer
155  rate be very low.
156
157  The recommended solution in such a case is to disable TSO/LSO, at
158  least on the relevant port. See for example:
159  https://atomicit.ca/kb/articles/slow-network-speed-windows-10/. If
160  that doesn’t help then "Don't Fragment" TCP flag may be
161  disabled. See for example:
162  https://support.microsoft.com/en-us/help/900926/recommended-tcp-ip-settings-for-wan-links-with-a-mtu-size-of-less-than. However,
163  note that disabling the “Don’t Fragment” flag may cause other
164  issues.
165
166  To test whether TSO/LSO may be the problem, do the following:
167
168  * If different machine configurations are used for the client and
169    server, try the iperf3 reverse mode (``-R``). If TSO/LSO is only
170    enabled on the client machine, this test should succeed.
171  * Reduce the sending length to a small value that should not require
172    segmentation, using the iperf3 ``-l`` option, e.g. ``-l 512``. It
173    may also help to reduce the MTU by using the iperf3 ``-M`` option,
174    e.g. ``-M 1460``.
175  * Using tools like Wireshark, identify the required MSS in the ICMP
176    Fragmentation Needed messages (if reported). Run tests with the
177    ``-l`` value set to 2 times the MSS and then 4 times, 6 times,
178    etc. With TSO/LSO issue in each test the throughput should be
179    reduced more. It may help to increase the testing time beyond the
180    default 10 seconds to better see the behavior (iperf3 ``-t``
181    option).
182
183What congestion control algorithms are supported?
184  On Linux, run this command to see the available congestion control
185  algorithms (note that some algorithms are packaged as kernel
186  modules, which must be loaded before they can be used)::
187
188    /sbin/sysctl net.ipv4.tcp_available_congestion_control
189
190  On FreeBSD, the equivalent command is::
191
192    /sbin/sysctl net.inet.tcp.cc.available
193
194I’m using the ``--logfile`` option. How do I see file output in real time?
195  Use the ``--forceflush`` flag.
196
197I'm using the --fq-rate flag, but it does not seem to be working. Why?
198  You need to add 'net.core.default_qdisc = fq' to /etc/sysctl.conf for that option to work.
199
200I'm having trouble getting iperf3 to work on Windows, Android, etc. Where can I get help?
201  iperf3 only supports Linux, FreeBSD, and OSX. For other platforms we recommend using iperf2.
202
203I managed to get a Windows executable built, but why do I get a BSOD on Windows 7?
204  There seems to be a bug in Windows 7 where running iperf3 from a
205  network filesystem can cause a system crash (in other words Blue
206  Screen of Death, or BSOD).  This is a Windows bug addressed in kb2839149:
207
208  https://support.microsoft.com/en-us/help/2839149/stop-error-0x00000027-in-the-rdbss-sys-process-in-windows-7-or-windows
209
210  A hotfix is available under kb2732673:
211
212  https://support.microsoft.com/en-us/help/2732673/-delayed-write-failed-error-message-when--pst-files-are-stored-on-a-ne
213
214Why can’t I run a UDP client with no server?
215  This is potentially dangerous, and an attacker could use this for a
216  denial of service attack.  We don't want iperf3 to be an attack tool.
217
218I'm trying to use iperf3 to test a 40G/100G link...What do I need to know?
219  See the following pages on fasterdata.es.net:
220
221  - https://fasterdata.es.net/host-tuning/100g-tuning/
222  - https://fasterdata.es.net/performance-testing/network-troubleshooting-tools/iperf/multi-stream-iperf3/
223
224My receiver didn't get all the bytes that got sent but there was no loss.  Huh?
225  iperf3 uses a control connection between the client and server to
226  manage the start and end of each test.  Sometimes the commands on
227  the control connection can be received and acted upon before all of
228  the test data has been processed.  Thus the test ends with data
229  still in flight.  This effect can be significant for short (a few
230  seconds) tests, but is probably negligible for longer tests.
231
232A file sent using the ``-F`` option got corrupted...what happened?
233  The ``-F`` option to iperf3 is not a file transfer utility.  It's a
234  way of testing the end-to-end performance of a file transfer,
235  including filesystem and disk overheads.  So while the test will
236  mimic an actual file transfer, the data stored to disk may not be
237  the same as what was sent.  In particular, the file size will be
238  rounded up to the next larger multiple of the transfer block size,
239  and for UDP tests, iperf's metadata (containing timestamps and
240  sequence numbers) will overwrite the start of every UDP packet
241  payload.
242
243I have a question regarding iperf3...what's the best way to get help?
244  Searching on the Internet is a good first step.
245  http://stackoverflow.com/ has a number of iperf3-related questions
246  and answers, but a simple query into your favorite search engine can
247  also yield some results.
248
249  There is a mailing list nominally used for iperf3 development,
250  iperf-dev@googlegroups.com.
251
252  We discourage the use of the iperf3 issue tracker on GitHub for
253  support questions.  Actual bug reports, enhancement requests, or
254  pull requests are encouraged, however.
255
256
257