Transperf

Transperf is a tool for testing transport protocol performance. Its key features include:

• Testing TCP performance over a variety of emulated network scenarios (using netem), including RTT, bottleneck bandwidth, and policed rate that can change over time.
• Testing multiple flows, on multiple machines.
• Testing multiple congestion control algorithms in a single test.
• Using a Python-based language for specifying test configurations and making assertions about the expected results.
• Visualizing results using a web GUI, along with archiving the configuration, logs, pcap traces, tcptrace xplot files, and outputs of a test.

Note that this is not an officially supported Google product.

CHANGELOG

Please see the CHANGELOG for a release history.

Documentation quick links

• Installation
• User Guide
• Testing Metrics
• Examples
• Frequently Asked Questions

Installation

Because transperf is launched from your local machine and run on a remote host, there are steps that need to be taken on both your local machine and the remote host.

Setup Remote Hosts

The following instructions are specifically for running transperf using GCE VMs using Debian 9, and may need to be adjusted depending on your testing environment. The number of instances you need to create depends on the types of tests you want to run. At minimum, you need two: one receiver and one sender instance.

1. For each instance, first ensure that each remote host is up-to-date and has these dependencies installed.

   On the remote hosts:

   $ sudo apt update && sudo apt upgrade
   $ sudo apt install rsync psmisc ifenslave pkg-config bison flex git make \
    	      bridge-utils util-linux tcpdump
   

   Some distributions have an out-of-date version of iproute2, so you should install the latest version from source.

   $ git clone git://git.kernel.org/pub/scm/network/iproute2/iproute2.git
   $ cd iproute2/
   $ ./configure
   $ sudo make install
   

2. Then, if you want to test bbrv2, follow the instructions on the bbr repository to install a kernel with bbrv2 on each of the remote hosts.

3. Configure rsyslog on remote hosts. By default, rsyslog on Google Cloud collects logs too aggressively, which causes issues in collecting kernel log output needed by transperf. On each remote machine, edit the following file

   $ sudo vim /etc/rsyslog.d/90-google.conf
   

   by commenting out the final line, as shown below.

   # Google Compute Engine default console logging.
   #
   # daemon: logging from Google provided daemons.
   # kern: logging information in case of an unexpected crash during boot.
   #
   # daemon,kern.* /dev/console

   Then, restart rsyslog:

   $ sudo service rsyslog restart
   

4. Enable root login. Transperf needs to execute as root on all remote machines, so root login must be enabled by editing the following file

   $ sudo vim /etc/ssh/sshd_config
   

   by finding the line containing PermitRootLogin and setting it to yes:

   ...
   # Authentication:
   
   #LoginGraceTime 2m
   PermitRootLogin yes
   #StrictModes yes
   #MaxAuthTries 6
   #MaxSessions 10
   ...

   Then, restart sshd:

   $ sudo service sshd restart
   

   Verify that this worked by ensuring you can ssh into the remote machine as root:

   $ ssh root@$HOST
   

   Note that you may need to use gcloud to establish SSH keys once. However, after they keys are exchanged, you should be able to ssh directly.

5. Setup a bonding device on the receiver. Whichever remote host you want to use as your receiver must be set up with a bonding device named eth0, and the physical NIC named eth1. To do this, rename the physical device from the default value (e.g. eth0 or ens4) to eth1 by adding the following line to the specified file.

   /etc/udev/rules.d/70-persistent-net.rules:

   SUBSYSTEM=="net", ACTION=="add", ATTR{address}=="<nic hwaddr>", NAME="eth1"
   

   Then, create a bonding device named eth0, and have eth1 set as a slave.

   /etc/network/interfaces:

   # interfaces(5) file used by ifup(8) and ifdown(8)
   
   auto eth0
   
   iface eth0 inet dhcp
           bond-mode active-backup
           slaves eth1
   

   You will then need to reboot the VM for these changes to take effect. You can verify that it worked correctly by running

   $ ip address
   

   and checking that eth0 is a master device, and eth1 is a slave.

   On the senders, you will need to make sure the physical NIC is named eth0. If it is already named eth0 (e.g., on Debian) you do not need to do anything. Senders do not need a bonding device.

   /etc/udev/rules.d/70-persistent-net.rules:

   SUBSYSTEM=="net", ACTION=="add", ATTR{address}=="<nic hw addr>", NAME="eth0"
   

6. (If using GCE) Ensure that your firewall rules will not block tcp/udp traffic. Transperf uses ports 6200 and 6300 explicitly, and also uses randomly generated ports to transfer data between senders and receivers. Consequently, you need to make sure your firewall rules in GCP are set to allow tcp/udp traffic between your VMs. That is, if you run

   gcloud compute firewall-rules list
   

   you should see a line like

   NAME                              NETWORK  DIRECTION  PRIORITY  ALLOW            DENY  DISABLED
   ...
   default-xg4al7wofiqhpghuvl4aw6so  default  INGRESS    1000      tcp,udp                False
   ...
   

   if you have configured your firewall rules correctly.

Setup local machine

Your local machine must be the same architecture as your remote hosts (e.g., x86_64) because you will be compiling up-to-date binaries on your local machine, which will then be copied and run on the remote hosts. Your local machine is also responsible for processing all of the data retrieved from running the transperf experiment.

1. Install dependencies.

   $ sudo apt update
   $ sudo apt install python-scapy wireshark tcptrace autoconf texinfo \
                      git build-essential rsync
   

2. Build netperf from source to include burst and interval features.

   $ git clone -b netperf-2.7.0 https://github.com/HewlettPackard/netperf
   $ cd netperf
   $ ./autogen.sh
   $ ./configure --enable-burst --enable-intervals
   $ sudo make install
   

User Guide

Build

Unlike a typical Python package, transperf is run directly as a ZIP file. To prepare the ZIP from the root of the repository, just run:

$ make

Quick Start: Multi-server Mode

Most transperf experiments are run in multi-server mode. To run all the test configurations in the sample_config directory on host1, host2, and host3 by launching from your local machine, simply run:

$ python transperf.zip -v -x -c sample_config root@host1 root@host2 root@host3

If you're using GCE VMs, you will need to provide both their external IPs and internal IPs, like this:

$ python transperf.zip -v -x -c sample_config root@ext_ip1/int_ip1 root@ext_ip2/int_ip2 root@ext_ip3/int_ip3

Note that the first host provided as a command line argument is used as the receiver.

Quick Start: Single-server Mode

Transperf also has basic support for running tests using a single remote machine, rather than the normal workflow of requiring >= 2 remote machines. This is accomplished using Linux network containers (via netns). Note that TCP parameters are not virtualized. Consequently, tests in single-server mode are limited to using only a single set of parameters per congestion control algorithm.

To run a test configuration on the remote host1:

python transperf.zip -v -x -c sample_config/4bbr_50M_40ms_BDP.py --virtcleanup --ssrv host1 root@rcv root@snd1 root@snd2

Testing Metrics

Transperf includes the ability to set expectations for a variety of metrics and check the results of an experimental run against those expectations. This allows transperf to be used for high-level testing of transport protocol behavior.

These checks can be customized by including a dictionary of values, as shown below.

scores = {
    "tput": 90,  # use X% of total bw as goodput
    "rtt_med": 25,  # 100 * minRTT/rtt_i
    "lock_on_bw": 10,  # should exit STARTUP at bw at least x% of bottleneck
    "loss_avoid": 95,  # non-loss percentage
    "fairness": 75,  # jain fairness index * 100 over lifetime
    "convergence_fairness": 85,  # jain fairness for each second after conv
    "convergence_sec": 1.1,  # how long before asserting convergence
}
check = check_with_scores(scores)

Each of these values corresponds to one of the following metrics.

Goodput (tput)

Throughput represents the bits transmitted per second measured during the test for all the connections, while goodput represents transmission excluding retransmission and header overhead. The tput metric reports the total goodput. We normalize tput into range [0,100] by dividing the total goodput, b, by the bottleneck bandwidth, BW, and multiplying by 100.

tput = 100 * b / BW

RTT Reduction (rtt_med)

Round-trip time (RTT) represents the two-way latency measured by the congestion control module. To assign a score that is higher with lower RTT, we use the term RTT reduction. We normalize the RTT reduction as follows:

rtt_med = 100 * minRTT_i / medRTT_i

where minRTT and medRTT are the minimum and median measured RTTs for a flow, i.

Lock on Bandwidth (lock_on_bw)

This metric represents the ratio of the bandwidth at which BBR exists STARTUP. That is, it checks that

measured_lock_on_bw_i >= lock_on_target

where

lock_on_target = (lock_on_bw / 100) * btl_bw / num_conns

and subscript i represents that this condition is checked for each flow i.

Loss Avoidance (loss_avoid)

Given a loss rate, l, ranging from 0 to 1, the normalized loss avoidance score is the non-loss percent:

loss_avoid = 100 * (1 - l)

This gives lower loss rate a higher score.

Fairness (fairness)

We use Jain's fairness index to represent how fair the goodput is allocated between connections during the test. We adjust the Jain's fairness index from (0,1] to (0,100] as follows.

f(b_1, b_2, ..., b_n) = 100 * (sum(b_i))^2 / (n * sum(b_i^2))

where n is the number of connections and b_i is the goodput for connection i.

Convergence Fairness (convergence_fairness)

Unlike Fairness, which is the pass/fail threshold for the Jain fairness index for each flow over its entire lifetime, convergence fairness is the minimum Jain's fairness index that the flow throughputs must have in every one-second time buck after t = convergence_sec.

Examples

Example configurations for transperf tests can be found in sample_config.

FAQs

I keep getting Permission denied (publickey) even though PermitRootLogin is yes

Try running ssh with verbose output (ssh -v). If you haven't connected to root using gcloud, do that once so that SSH keys can be created and transferred.

You may need to specify the identity file created by gcloud (e.g. ~/.ssh/google_compute_engine) on your local machine. We recommend just adding that key using ssh-agent:

$ eval `ssh-agent`
$ ssh-add ~/.ssh/google_compute_engine