498108ad70
The TCP packet size used on the passt L2 link (qemu socket) makes a huge difference to passt/pasta throughput; many of passt's overheads (chiefly syscalls) are per-packet. That packet size is largely determined by the MTU on the L2 link, so we benchmark for a number of different MTUs. That works well for the guest to host transfers. For the host to guest transfers, we purport to test for different MTUs, but we're not actually adjusting anything interesting. The host to guest transfers adjust the MTU on the "host's" (actually ns) loopback interface. However, that only affects the packet size for the socket going to passt, not the packet size for the L2 link that passt manages - passt can and will repack the stream into packets of its own size. Since the depacketization on that socket is handled by the kernel it doesn't have a lot of bearing on passt's performance. We can't fix this by changing the L2 link MTU from the guest side (as we do for guest to host), because that would only change the guest's view of the MTU, passt would still think it has the large MTU. We could test this by using the --mtu option to passt, but that would require restarting passt for each run, which is awkward in the current setup. So, for now, drop all the "small MTU" tests for host to guest. Signed-off-by: David Gibson <david@gibson.dropbear.id.au> Signed-off-by: Stefano Brivio <sbrivio@redhat.com> |
||
|---|---|---|
| .. | ||
| build | ||
| demo | ||
| distro | ||
| env | ||
| lib | ||
| memory | ||
| passt | ||
| passt_in_ns | ||
| pasta | ||
| pasta_options | ||
| pasta_podman | ||
| perf | ||
| two_guests | ||
| .gitignore | ||
| ci | ||
| find-arm64-firmware.sh | ||
| Makefile | ||
| nstool.c | ||
| passt.mbuto | ||
| passt.mem.mbuto | ||
| prepare-distro-img.sh | ||
| README.md | ||
| run | ||
| run_demo | ||
| valgrind.supp | ||
Scope
This directory contains test cases for passt and pasta and a simple POSIX shell-based framework to define them, and run them as a suite.
These tests can be run as part of a continuous integration workflow, and are also used to provide short usage demos, with video recording, for passt and pasta basic use cases.
Run
Dependencies
Packages
The tests require some package dependencies commonly available in Linux distributions. If some packages are not available, the test groups that need them will be selectively skipped.
This is a non-exhaustive list of packages that might not commonly be installed on a system, i.e. common utilities such as a shell are not included here.
Example for Debian, and possibly most Debian-based distributions:
build-essential git jq strace iperf3 qemu-system-x86 tmux sipcalc bats bc
catatonit clang-tidy cppcheck go isc-dhcp-common psmisc linux-cpupower socat
netcat-openbsd fakeroot lz4 lm-sensors qemu-system-arm qemu-system-ppc
qemu-system-misc qemu-system-x86 valgrind
NOTE: the tests need a qemu version >= 7.2, or one that contains commit 13c6be96618c ("net: stream: add unix socket"): this change introduces support for UNIX domain sockets as network device back-end, which qemu uses to connect to passt.
Other tools
Test measuring request-response and connect-request-response latencies use
neper, which is not commonly packaged by distributions and needs to be built
and installed manually:
git clone https://github.com/google/neper
cd neper; make
cp tcp_crr tcp_rr udp_rr /usr/local/bin
Virtual machine images are built during test executions using mbuto, the shell script is sourced via git as needed, so there's no need to actually install it.
Kernel parameters
Performance tests use iperf3 with rather large TCP receiving and sending windows, to decrease the likelihood of iperf3 itself becoming the bottleneck. These values need to be allowed by the kernel of the host running the tests. Example for /etc/sysctl.conf:
net.core.rmem_max = 134217728 net.core.wmem_max = 134217728
Further, the passt demo uses perf(1), relying on hardware events for performance counters, to display syscall overhead. The kernel needs to allow unprivileged users to access these events. Suggested entry for /etc/sysctl.conf:
kernel.perf_event_paranoid = -1
Special requirements for continuous integration and demo modes
Running the test suite as continuous integration or demo modes will record the terminal with the steps being executed, using asciinema(1), and create binary packages.
The following additional packages are commonly needed:
alien asciinema linux-perf tshark
Regular test
Just issue:
./run
from the test directory. Elevated privileges are not needed. Environment
variable settings: DEBUG=1 enables debugging messages, TRACE=1 enables tracing
(further debugging messages), PCAP=1 enables packet captures. Example:
PCAP=1 TRACE=1 ./run
Running selected tests
Rudimentary support to run a list of selected tests, without support for dependencies, is available. Tests need to have a setup function corresponding to their path. For example:
./run passt/ndp passt/dhcp pasta/ndp
will call the 'passt' setup function (from lib/setup), run the two corresponding tests, call the 'passt' teardown function, the 'pasta' setup, run the pasta/ndp test, and finally tear down the 'pasta' setup.
Note that requirements on steps implemented by related tests are not handled. For example, if the 'passt/tcp' needs guest connectivity set up by the 'passt/ndp' and 'passt/dhcp' tests, those need to be listed explicitly.
Continuous integration
Issuing:
./ci
will run the whole test suite while recording the execution, and it will also build JavaScript fragments used on http://passt.top/ for performance data tables and links to specific offsets in the captures.
Demo mode
Issuing:
./demo
will run the demo cases under demo, with terminal captures as well.
Framework
The implementation of the testing framework is under lib, and it provides
facilities for terminal and tmux session management, interpretation of test
directives, video recording, and suchlike. Test cases are organised in the
remaining directories.
Test cases can be implemented as POSIX shell scripts, or as a set of directives,
which are not formally documented here, but should be clear enough from the
existing cases. The entry point for interpretation of test directives is
implemented in lib/test.