OB-UDPST

Open Broadband-UDP Speed Test (OB-UDPST) is a client/server software utility to demonstrate one approach of doing IP capacity measurements as described by:

• Broadband Forum TR-471 Issue 3 (12/2022): Maximum IP-Layer Capacity Metric, Related Metrics, and Measurements, BBF TR-471, https://www.broadband-forum.org/technical/download/TR-471.pdf

• ITU-T Recommendation Y.1540 (revised 03/2023): Internet protocol data communication service - IP packet transfer and availability performance parameters, ITU-T Y.1540, https://www.itu.int/rec/T-REC-Y.1540-201912-I/en

• ITU-T Y-series Supplement 60 (2022): Interpreting ITU-T Y.1540 maximum IP-layer capacity measurements, ITU-T Y.Sup60, https://www.itu.int/rec/T-REC-Y.Sup60/en

• ETSI Technical Specification 103 222 Part 2 (2019-08): Reference benchmarking and KPIs for High speed internet, ETSI TS 103 222-2, V1.2.1, https://www.etsi.org/deliver/etsi_ts/103200_103299/10322202/01.02.01_60/ts_10322202v010201p.pdf

• IETF RFC 9097: Metrics and Methods for One-way IP Capacity, https://datatracker.ietf.org/doc/html/rfc9097

• ETSI Technical Report 103 702 (2020-11): Speech and multimedia Transmission Quality (STQ); QoS parameters and test scenarios for assessing network capabilities in 5G performance measurements, ETSI TR 103 702, V1.1.1 https://www.etsi.org/deliver/etsi_tr/103700_103799/103702/01.01.01_60/tr_103702v010101p.pdf

Overview

Utilizing an adaptive transmission rate, via a pre-built table of discreet sending rates (starting at 0.11 Mbps), UDP datagrams are sent from client to server(s) or server(s) to client to determine the maximum available IP-layer capacity between them. The load traffic is only sent in one direction at a time, and status feedback messages are sent periodically in the opposite direction.

For upstream tests, the feedback messages from the server(s) instruct the client on how it should adjust its transmission rate based on the presence or absence of sufficient sequence errors or delay variation changes observed by the server. For downstream tests, the feedback messages simply communicate any sequence errors or delay variation changes observed by the client. In either case, the server is the host executing the algorithm that determines the rate adjustments made during the test. This centralized approach allows the rate adjustment algorithm to be more easily enhanced and customized to accommodate diverse network services and protocols. To that end, and to encourage additional testing and experimentation, the software has been structured so that virtually all of the settings and thresholds used by the algorithm are currently available as client-side command-line parameters (allowing modification beyond the current set of default values).

By default, both IPv4 and IPv6 tests can be serviced simultaneously when acting as a server. When acting as a client, testing is performed in one address family at a time. Also, the default behavior is that jumbo datagram sizes (datagrams that would result in jumbo frames) are utilized for sending rates above 1 Gbps. The expectation is that jumbo frames can be accommodated end-to-end without fragmentation. For sending rates below 1 Gbps, or all sending rates when the -j option is used, the default UDP payload size is 1222 bytes. This size was chosen because although it is relatively large it should still avoid any unexpected fragmentation due to intermediate protocols or tunneling. Alternatively, the -T option is available to allow slightly larger default datagrams that would create full packets when a traditional 1500 byte MTU is available end-to-end. These larger sizes are optional because several consumer-grade routers have been shown to handle packet fragmentation, from a performance perspective, very poorly. Note that both the -j and -T options must match between the client and server or the server will reject the test request.

All delay variation values, based on One-Way Delay (OWDVar) or Round-Trip Time (RTTVar), indicate the delays measured above the most recent minimum. Although OWDVar is measured for each received datagram, RTTVar is only sampled and is measured using each status feedback message. If specifying the use of OWDVar instead of RTTVar (via the -o option), there is no requirement that the clocks between the client and server be synchronized. The only expectation is that the time offset between clocks remains nominally constant during the test.

Usage examples for server mode:

$ udpst
    Service client requests received on any interface
$ udpst -x
    Service client requests in background (as a daemon) on any interface
$ udpst -p <port> <Local_IP>
    Service client requests using a non-default UDP control port and only when
    received on the interface with the specified IP address
$ udpst -a <key>
    Only service requests that utilize a matching authentication key

Note: The server must be reachable on the UDP control port [default 25000] and all UDP ephemeral ports (32768 - 60999 as of the Linux 2.4 kernel, available via cat /proc/sys/net/ipv4/ip_local_port_range).

Usage examples for client mode:

$ udpst -u <server>
    Do upstream test from client to server using 1 connection (UDP flow)
$ udpst -d <server> <server> <server>
    Do downstream test from 3 server instances using 1 connection each
$ udpst -C 4 -p <port> -d <server>
    Do downstream test from 1 server instance using 4 connections and a
    non-default UDP control port
$ udpst -C 3 -u <server> <server> <server> <server>
    Do upstream test to 4 server instances using 4 connections, but only
    requiring a minimum of 3 (one server could be unreachable or unavailable)
$ udpst -C 6 -d <server>:<port> <server>:<port> <server>:<port>
    Do downstream test from 3 server instances, each using a non-default UDP
    control port, with 6 total connections (setup round robin, 2 each)
$ udpst -C 6-8 -u <server> <server> <server> <server>
    Do upstream test to 4 server instances using 8 connections, but only
    requiring a minimum of 6 (one server could be unreachable or unavailable)
$ udpst -r -d <server>
    Do downstream test and show loss ratio instead of delivered percentage
$ udpst -f json -d <server> >udpst.json
    Do downstream test and redirect JSON output to a file

Note: The client can operate behind a firewall (w/NAT) without any port forwarding or DMZ designation because all connections with the server(s) are initiated by the client.

To show the sending rate table:

$ udpst -S

The table shows dual transmitters for each index (row) because two separate sets of transmission parameters are required to achieve the desired granularity of sending rates. Each of the transmitters has its own interval timer and one also includes an add-on fragment at the end of each burst (again, for granularity). By default the table is shown for IPv4 with jumbo datagram sizes enabled above 1 Gbps. Including -6 (for IPv6 only), -j (to disable all jumbo sizes), or -T (for traditional 1500 byte MTU sizes) shows the table adjusted for those specific scenarios.

For a list of all options:

$ udpst -?

Default values:

Note that default values have been provided for all essential settings and thresholds, and these values are recommended for use at the beginning of any test campaign. The set of default values are subject to re-consideration with further testing, and may be revised in the future.

There are circumstances when changes to the defaults are warranted, such as extremely long paths with unknown cross traffic, high levels of competing traffic, testing over radio links with highly variable loss and delay, and test paths that exhibit bi-modal rate behavior in repeated tests.

An option in Release 7.5.0 allows the client to request the algorithm used for load adjustment when conducting a search for the Maximum IP-Layer Capacity. The Type C algorithm (a.k.a. Multiply and Retry) will provide a fast rate increase until congestion, reaching 1 Gbps in ~1 second. The "fast" ramp-up will be re-tried when conditions warrant, to ensure that the Maximum IP-Layer Capacity has been reached. This option is activated using -A c (with the more linear Type B algorithm remaining the default).

One change to the default settings was included in Release 7.5.1. All Load Adjustment search algorithms will now Ignore Reordering (and duplication) as a component of sequence errors: only packet loss will increase the sequence error count. The optional use of the -R option will now revert the behavior from "Ignore" back to "Include". This change was justified by recognizing that both out-of-order and duplicate datagrams are a legitimate part of IP-Layer Capacity.

See the following publication (which is updated frequently) for more details on testing in the circumstances described above:

• ITU-T Y-series Supplement 60 (06/22): Interpreting ITU-T Y.1540 maximum IP-layer capacity measurements, ITU-T Y.Sup60, https://www.itu.int/rec/T-REC-Y.Sup60/en

Multiple Connections and Distributed Servers

As of Release 8.0.0, the client can now test using multiple connections (i.e., UDP flows) to one or more server instances. Each server instance can itself service up to 256 independent client connections. When the client wants to establish more than one connection per server instance OR the client wants to specify a minimum (and optional maximum) number of connections, the -C cnt[-max] option is used.

For better utilization of hosts, multiple server instances can reside on one physical machine and service test requests across one or more network interfaces using unique IP addresses (including IP aliases) or different UDP control ports. For redundancy and load balancing (within or between locations), multiple server instances can be utilized across completely separate physical servers, virtual machines, or containers. And because OB-UDPST is based on UDP, it is significantly less sensitive to delay than TCP-based measurement tools (allowing much greater flexibility with geographic server placement).

Some of the benefits of testing with multiple connections to multiple server instances include:

• More efficient use of server resources (CPU cores, network interfaces, etc.)
• Better utilization of any link aggregation within the network
• Increased use of ECMP (Equal Cost Multi-Path) in the data center or WAN
• The ability to ramp-up test traffic at a compounded rate
• Improved capacity to deal with competing traffic, particularly with Active Queue Management (AQM) schemes that provide flow queuing/isolation
• The option to utilize smaller servers or lower-speed interfaces instead of a few large machines with high-speed connectivity
• Support for maintenance and downtime on individual servers without test interruption
• Opportunities for various levels of redundancy by way of diverse server-side resources (i.e., NICs, physical machines, LAN networks, WAN links, entire data centers,...)

Important Considerations

There will be different system and network resource impacts for the client and server given the contrasting downstream vs. upstream (N-to-1 vs. 1-to-N) traffic dynamic. Very often, more connections will not equal more capacity (the law of diminishing returns is ever-present). And while low-speed testing up to 1 Gbps can typically utilize more connections (10+) effectively, high-speed testing above that threshold generally performs best with fewer connections (2-4).

With high-speed testing the number of connections can significantly impact achievable rates due to the default udpst behavior of using jumbo size datagrams for sending rates above 1 Gbps (i.e., when -j is not used). With fewer connections, where each one will need to drive traffic above 1 Gbps, they will be able to take advantage of the much higher network efficiency and reduced I/O rate of larger datagrams. Of course, this assumes that jumbo frames are supported by the network and IP fragmentation can be avoided.

In another situation observed on a Raspberry Pi 4 (but called out here for its possible relevance to other devices), the use of multiple flows caused an undesirable outbound congestion condition on the local interface. The result was unreliable and inconsistent measurements because some flows would either monopolize all the bandwidth or be completely starved of it. The cause of the issue was its multi-queue network interface coupled with its default behavior of assigning flows to those queues using a 4-tuple hash. Depending on how the flows were hashed, some were able to completely starve the others. And this behavior was not unique to udpst. The same multi-flow starvation was observed with iPerf, for both UDP and TCP.

One simple way to avoid this issue on the Pi4 is to do testing with a single connection. Alternatively, XPS (Transmit Packet Steering) can be utilized to make the device behave the same as servers normally do with multi-queue NICs. That is, having each CPU map onto one queue. Additional details, as well as the XPS configuration used on the Pi4, are available below under "Considerations for Older Hardware and Low-End Devices".

Building OB-UDPST

To build OB-UDPST a local installation of CMake is required. Please obtain it for your particular build system either using the locally available packages or consult with [https://cmake.org] for other download options.

$ cmake .
$ make

Note: Authentication functionality uses a command-line key along with the OpenSSL crypto library to create and validate a HMAC-SHA256 signature (which is used in the setup request to the server). Although the makefile will build even if the expected directory is not present, disabling the key and library dependency, the additional files needed to support authentication should be relatively easy to obtain (e.g., sudo apt-get install libssl-dev or sudo yum install openssl-devel).

Test Processing Walkthrough (all messaging and PDUs use UDP)

On the server, the software is run in server mode in either the foreground or background (as a daemon) where it awaits Setup requests on its UDP control port.

The client, which always runs in the foreground, requires a direction parameter as well as the hostname or IP address of one or more servers. The client will create a connection and send a Setup request to each server's control port. It will also start a test initiation timer for each so that if the initiation process fails to complete, and the required minimum connection count is not available, the client will display an error message to the user and exit.

Setup Request

When the server receives the Setup request it will validate the request by checking the protocol version, the jumbo datagram support indicator, and the authentication data if utilized. If the Setup request must be rejected, a Setup response will be sent back to the client with a corresponding command response value indicating the reason for the rejection. If the Setup request is accepted, a new test connection is allocated and initialized for the client. This new connection is associated with a new UDP socket allocated from the UDP ephemeral port range. A timer is then set for the new connection as a watchdog (in case the client goes quiet) and a Setup response is sent back to the client. The Setup response includes the new port number associated with the new test connection. Subsequently, if a Test Activation request is not received from the client on this new port number, the watchdog will close the socket and deallocate the connection.

Setup Response

When the client receives the Setup response from the server it first checks the command response value. If it indicates an error it will display a message to the user (and exit if the required connection count falls below the minimum). If it indicates success it will build a Test Activation request with all the test parameters it desires such as the direction, the duration, etc. It will then send the Test Activation request to the UDP port number the server communicated in the Setup response.

Test Activation request

After the server receives the Test Activation request on the new connection, it can choose to accept, ignore or modify any of the test parameters. When the Test Activation response is sent back, it will include all the test parameters again to make the client aware of any changes. If an upstream test is being requested, the transmission parameters from the appropriate row of the sending rate table are also included. Note that the server additionally has the option of completely rejecting the request and sending back an appropriate command response value. If activation continues, the new connection is prepared for an upstream OR downstream test with either a single timer to send status PDUs at the specified interval OR dual timers to send load PDUs based on the specific row of the sending rate table. The server then sends a Test Activation response back to the client, the watchdog timer is updated and a test duration timer is set to eventually stop the test. The new connection is now ready for testing.

Test Activation response

When the Test Activation response is received back at the client it first checks the command response value. If it indicates an error it will display a message to the user (and exit if the required connection count falls below the minimum). If it indicates success it will update any test parameters modified by the server. It will then prepare its connection for an upstream OR downstream test with either dual timers set to send load PDUs (based on the transmission parameters sent by the server) OR a single timer to send status PDUs at the specified interval. The test initiation timer is then stopped and a watchdog timer is started (in case the server goes quiet). The connection is now ready for testing.

Testing

Testing proceeds with one end point sending load PDUs, based on transmission parameters from the sending rate table, and the other end point sending status messages to communicate the traffic conditions at the receiver. Each time a PDU is received the watchdog timer is reset. When the server is sending load PDUs it is using the transmission parameters directly from the sending rate table via the index that is currently selected (which was based on the feedback in its received status messages). However, when the client is sending load PDUs it is not referencing a sending rate table but is instead using the discreet transmission parameters that were communicated by the server in its periodic status messages. This approach allows the server to always control the individual sending rates as well as the algorithm used to decide when and how to adjust them.

Test Stop

When the test duration timer on the server expires it sets the connection test action to STOP and also starts marking all outgoing load or status PDUs with a test action of STOP. When received by the client, this is the indication that it should finalize testing, display the test results, and also mark its connection with a test action of STOP (so that any subsequently received PDUs are ignored). With the test action of the connection set to STOP, the very next expiry of a send timer for either a load or status PDU will cause the client to schedule an immediate end time to exit. It then sends that PDU with a test action of STOP as confirmation to the server. When the server receives this confirmation in the load or status PDU, it schedules an immediate end time for the connection which closes the socket and deallocates it.

More Info

For much more detail on the test protocol, see the ./protocol.md file. Also, an Internet-Draft, https://datatracker.ietf.org/doc/draft-ietf-ippm-capacity-protocol/ describes Protocol Version 9 in even more detail.

JSON Output

For examples of the JSON output fields see the included sample files named "udpst-*.json". Available JSON output options include -f json (unformatted), -f jsonb (brief & unformatted), and -f jsonf (formatted). To significantly reduce the size of the JSON output, option -s (omit sub-interval results) can be combined with -f jsonb (omit static input fields).

Included in the output is a numeric ErrorStatus field (which corresponds with the software exit status) as well as a text ErrorMessage field. If a test completes normally without incident the error status will be 0 (zero) and the error message will be empty. If a test completes, but encounters a soft error or warning, the error status will be 1 (one) and the most recent warning message will be included. If a test fails to complete, the error status will be -1 (negative one) and the most recent warning or error message will be included.

The file "ob-udpst_output_mapping.pdf" provides a mapping between JSON key names, TR-471 names, TR-181 names, and the ob-udpst STDOUT names for various results.

Note: When stdout is not redirected to a file, JSON may appear clipped due to non-blocking console writes.

Local Interface Traffic Rate

Where applicable, it is possible to also output the local interface traffic rate (in Mbps) via the -E intf option. This can be informative when trying to account for external traffic that may be consuming a non-trivial amount of the interface bandwidth and competing with the measurement traffic. The rate is obtained by querying the specific interface byte counters that correspond with the direction of the test (i.e., tx_bytes for upstream tests and rx_bytes for downstream tests). These values are obtained from the sysfs path /sys/class/net/<intf>/statistics. An additional associated option -M is also available to override normal behavior and use the interface rate instead of the measurement traffic to determine a maximum.

When the -E intf option is utilized, the console output will show the interface name in square brackets in the header info and the Ethernet rate of the interface in square brackets after the L3/IP measured rate. When JSON output is also enabled, the interface name appears in "Interface" and the interface rate is in "InterfaceEthMbps". When this option is not utilized, these JSON fields will contain an empty string and zero respectively.

Server Bandwidth Management

The -B mbps option can be used on a server to designate a maximum available bandwidth. Often, this would simply specify the speed of the interface servicing tests and is managed separately for upstream and downstream (i.e., -B 1000 indicates that 1 Gbps is available in each direction). One scenario to achieve better server utilization is to run multiple server instances on a machine with them bound to one or more different physical interfaces. In this scenario, the bandwidth option for each is set to handle some portion of the total aggregate available (with clients always utilizing several at once). For example, 10 instances could be run with -B 1000 for a 10G or -B 10000 for a 100G. This can be accomplished by either binding each to a different IP alias or configuring them to use different UDP control ports (e.g., -p 25000, -p 25001, -p 25002,...). And although single threaded, each running instance supports multiple simultaneous overlapping tests.

When configured on the server, clients will also need to utilize the -B mbps option in their test request to indicate the maximum bandwidth they may require from the server for accurate testing. For example, a client connected via a 300 Mbps Internet service would specify -B 300. If the server's available bandwidth is unable to accommodate what the client requires, due to tests already in progress, the test request is rejected and the client produces an error indicating this as the cause. The client can then retry the test a little later (when server bandwidth may be available) or immediately try either an alternate server instance (per the scenarios described above) or a different server altogether. When the client is testing with multiple connections, the provided bandwidth option is evenly divided across the attempted connections. In this case, if only some connections are rejected due to insufficient server capacity, and the required minimum connection count is available, the testing will proceed normally.

Note: This option does not alter the test methodology or the rate adjustment algorithm in any way. It only provides test admission control to better manage the server's network bandwidth.

Increasing the Starting Sending Rate (Considerations)

While the -I index option designates a fixed sending rate, it is also possible to set the starting rate with load adjustment enabled. Option -I @index allows selection of a higher initial sending rate starting at the specified index in the sending rate table (with the default equivalent to -I @0 and shown as <Auto>). Note that with or without the @ character prefix, this option relies on an index from the sending rate table (see -S for index values).

For maximum capacities up to 1 Gbps, the -h delta option has been available for some time to allow customization of the ramp-up speed in the Type B Algorithm. In some cases, and especially with few connections and maximum capacities above 1 Gbps (where -h delta no longer has an impact), it can make sense to start at an initial sending rate above index 0 (zero). For example, if testing a 10 Gbps service with only one connection, specifying -I @1000 would start with a sending rate of 1 Gbps.

One important consideration with the -I @index option is that setting too high a value can be counter-productive to finding an accurate maximum. This is because when some devices or communication channels are suddenly overwhelmed by the appearance of very-high sustained traffic, it can result in early congestion and data loss that make it prematurely appear as if a maximum capacity has been reached. As such, it is recommended to use starting rates of only 10-20% of the expected maximum to avoid an early overload condition and false maximum.

Linux Socket Buffer Optimization

For high speed testing (typically above 1 Gbps), the socket buffer maximums of the Linux kernel can be increased to reduce possible datagram loss. As an example, the following could be added to the /etc/sysctl.conf file:

net.core.rmem_max=16777216
net.core.wmem_max=16777216

To activate the new values without a reboot:

$ sudo sysctl -w net.core.rmem_max=16777216
$ sudo sysctl -w net.core.wmem_max=16777216

The default software settings will automatically take advantage of the increased send and receive socket buffering available (shown in verbose mode). However, the command-line option -b buffer can be used if even higher buffer levels (granted at 2x the designated value) should be explicitly requested for each socket.

Server Optimization - Particularly When Jumbo Frames Are Unavailable

By default, the sending rate table utilizes jumbo size datagrams when testing above 1 Gbps. As expected, maximum performance is obtained when the network also supports a jumbo MTU size (9000+ bytes). However, some environments are restricted to a traditional MTU of 1500 bytes and would be required to fragment the jumbo datagrams into multiple IP packets.

In these situations, the recommendation is to utilize the -j option on both the client and server to restrict all datagrams to non-jumbo sizes. However, because of the resulting higher socket I/O rate at high speeds, this may limit the maximum rate that can be achieved. If jumbo size datagrams are still desired and udpst was compiled with the GSO (Generic Segmentation Offload) optimization, the default with reasonably recent Linux kernels, it will need to be recompiled without it as GSO is incompatible with IP fragmentation. This can be accomplished via the following:

$ cmake -D HAVE_GSO=OFF .

NUMA Node Selection

An important performance consideration is to instantiate the udpst processes in the same Non-Uniform Memory Access (NUMA) node as the network interface. This placement will limit cross-NUMA memory access on systems with more than one NUMA node. Testing has shown this commonplace server optimization can significantly increase sending rates while also reducing CPU utilization. The goal is to set the CPU affinity of the udpst process to the same NUMA node as the one handling the network interface used for testing. Follow the steps below to achieve this:

First, obtain the NUMA node count (to verify applicability if >1) as well as the CPU listing for each node.

$ lscpu | grep NUMA
NUMA node(s):                    2
NUMA node0 CPU(s):               0-13,28-41
NUMA node1 CPU(s):               14-27,42-55

Next, find the NUMA node that corresponds to the test interface (in this case ens1f0 is associated with node 0).

$ cat /sys/class/net/ens1f0/device/numa_node
0

Finally, start the server instances with a CPU affinity that matches the NUMA node of the test interface (node0 = 0-13,28-41).

By IP address:
$ taskset -c 0-13,28-41 udpst -x <Local_IP1>
$ taskset -c 0-13,28-41 udpst -x <Local_IP2>
...
$ taskset -c 0-13,28-41 udpst -x <Local_IPN>

By UDP control port:
$ taskset -c 0-13,28-41 udpst -x -p <Port1> <Local_IP>
$ taskset -c 0-13,28-41 udpst -x -p <Port2> <Local_IP>
...
$ taskset -c 0-13,28-41 udpst -x -p <PortN> <Local_IP>

In addition to the udpst server instances, multi-queue NICs will also heavily utilize various CPUs on the same NUMA node for interrupt handling and I/O. This contention for CPU resources should be considered when determining how many udpst instances to run. In some circumstances it might be beneficial to differentiate the CPU cores based on usage. One way to do this involves banning some CPUs from interrupt handling in the irqbalance environment file via the "IRQBALANCE_BANNED_CPULIST" variable (older versions use a "IRQBALANCE_BANNED_CPUS" mask instead). Those CPUs that would then no longer be burdened with also processing interrupts could be the ones specified in the taskset CPU list when running udpst instances.

The example discussed here only addresses a single network interface on one NUMA node. Server utilization in this case could be further maximized by also setting up a second network interface on the other NUMA node and repeating the appropriate configuration. Ideally, always growing the server by two interfaces at a time (one on each node).

Fragment Reassembly Memory

If the -j option is not used and IP fragmentation of jumbo size datagrams must be expected as a normal part of testing (where udpst must also be built without the GSO optimization), it is important to make sure that adequate memory is available for fragment reassembly. When not available, the "packet reassemblies failed" counter under netstat -s and/or netstat -s -6 (for IPv6) will show the failures.

To increase the memory available to reassemble fragments, as well as limit the time a fragment should be kept awaiting reassembly, the following could be added to the /etc/sysctl.conf file:

net.ipv4.ipfrag_high_thresh=104857600
net.ipv4.ipfrag_time=3
net.ipv6.ip6frag_high_thresh=104857600
net.ipv6.ip6frag_time=3

To activate the new values without a reboot:

$ sudo sysctl -w net.ipv4.ipfrag_high_thresh=104857600
$ sudo sysctl -w net.ipv4.ipfrag_time=3
$ sudo sysctl -w net.ipv6.ip6frag_high_thresh=104857600
$ sudo sysctl -w net.ipv6.ip6frag_time=3

The suggested thresholds shown above are 25x the typical default. However, if the "packet reassemblies failed" counter continues to increase during testing, the threshold values should be raised accordingly.

Transmit and Receive Rings

For higher speed NICs (10G and above), increasing the transmit and receive ring size is often required to maximize system throughput. To see the current value and supported maximums, do the following:

$ ethtool -g <intf>

To increase the ring sizes to their available maximums, do the following:

$ sudo ethtool -G <intf> rx <max> tx <max>

The settings will need to be added to the system configuration for the new values to persist across reboots. The details of formally incorporating ethtool settings into the boot process are distribution specific. However, an "informal" approach could simply make use of the /etc/rc.local file.

Considerations for Older Hardware and Low-End Devices

There are two general categories of devices in this area, 1) those that operate normally but lack the horsepower needed to reach a specific sending rate and 2) those that are unable to function properly because they do not support the required clock resolution. In this case, an error message is produced (“ERROR: Clock resolution (xxxxxxx ns) out of range”) and the software exits because without the expected interval timer, sending rates would be skewed and the rate-adaption algorithm would not function properly.

Devices in the first category may be helped by the -T option when a traditional 1500 byte MTU is available end-to-end between client and server. This option will increase the default datagram payload size so that full 1500 byte packets are generated. This reduces both the socket I/O and network packet rates (see -S vs. -ST output). Another important option in these cases, when jumbo frames are not supported by the network, is -j. This will disable all jumbo datagram sizes and prevent any possible IP fragmentation. This can happen with a very underpowered device when the server, attempting to drive the client higher and higher, ends up at sending rates above 1 Gbps.

One specific device in this category worth mentioning is a Raspberry Pi 4 running Raspberry Pi OS (previously called Raspbian). Testing has shown that to reach a 1 Gbps sending rate, it was necessary to set the CPU affinity of udpst to avoid CPU 0 (the CPU handling network interrupts when irqbalance is not used). Additionally, and especially when using a 32-bit Raspbian, both the -T and -j options were also needed. And these options are recommended for any device in this general category. The command to utilize these recommendations is:

$ taskset -c 1-3 udpst -u -T -j <server>

Before moving on, a final consideration for the Raspberry Pi 4 (because this may apply to other devices) has to do with multi-connection testing. Whenever multiple flows aggressively congest the outbound multi-queue network interface, the default behavior of 4-tuple hashing for queue assignment will allow some flows to monopolize all the available bandwidth while others will be starved of it. This can cause reliability and consistency issues with measurements as some flows experience timeouts, or shutdown completely. Other than limiting tests to only a single connection, XPS (Transmit Packet Steering) can be utilized to align the Pi4 behavior with a typical server using multi-queue NICs (i.e., each CPU mapped onto one queue). To achieve this for eth0, the following can be added to the /etc/rc.local file:

echo 1 > /sys/class/net/eth0/queues/tx-0/xps_cpus
echo 2 > /sys/class/net/eth0/queues/tx-1/xps_cpus
echo 4 > /sys/class/net/eth0/queues/tx-2/xps_cpus
echo 8 > /sys/class/net/eth0/queues/tx-3/xps_cpus

For devices in the second category mentioned above (unsupported timer resolution), a compile-time option (DISABLE_INT_TIMER) is available that does not rely on an underlying system interval timer. However, the trade-off for this mode of operation is that it results in high CPU utilization. But, clients running on older or low-capability hosts may be able to execute tests where they otherwise would not.

$ cmake -D DISABLE_INT_TIMER=ON .

Note: Because of the increased CPU utilization, this option is not recommended for standard server operation.