]> Széchenyi István University

Egyetem tér 1. Győr H-9026 Hungary lencse@sze.hu

SoftBank Corp.

1-7-1 Kaigan Minato-ku Tokyo 105-7529 Japan shima@wide.ad.jp https://softbank.co.jp/

Operations and Management Area Benchmarking Methodology Working Group Benchmarking, Measurement Procedure, Receive-Side Scaling, Multiple IP Addresses RFC 2544 has defined a benchmarking methodology for network interconnect devices. Its test frame format contained fixed IP addresses and fixed port numbers. RFC 4814 introduced pseudorandom port numbers but used a single source and destination IP address pair when testing with a single destination network. This limitation may cause an issue when the device under test uses the Receive-Side Scaling (RSS) mechanism in the packet processing flow. RSS has two implementations: the first only includes the IP addresses, whereas the second also includes the port numbers in the tuple used for hashing. Benchmarking tests that use a single IP address pair and RFC 4814 pseudorandom port numbers are biased against the first type of RSS implementation because traffic is not distributed among the processing elements. This document recommends the usage of pseudorandom IP addresses in a similar manner as RFC 4814 did with the port numbers. If accepted, this document updates all affected RFCs, including RFC 2544, RFC 4814, RFC 5180, RFC 8219.

has defined a comprehensive benchmarking methodology for network interconnect devices, which is still in use. It was amended by several RFCs, which did not formally update it. introduced pseudorandom port numbers (instead of fixed ones). addressed IPv6 specificities and added technology updates but declared IPv6 transition technologies out of scope. addressed the IPv6 transition technologies. Receive-Side Scaling (RSS) aims to distribute the workload caused by packet arrivals among the CPU cores evenly to achieve high performance. It has two types of implementations: the first one only includes the IP addresses, whereas the second one also includes the port numbers into the tuple used for hashing . compliant testers work properly in the second case; however, the pseudorandom port numbers cannot provide entropy if the Device Under Test (DUT) follows the first type of RSS implementation; therefore, these devices produce poor benchmarking results in compliant laboratory tests, whereas they can exhibit high performance in production environments where the usage of multiple IP addresses ensures the entropy for the proper operation of their RSS implementation. Therefore, the conditions of laboratory tests should be improved to ensure unbiased performance testing. To that end, this document examines how the usage of multiple IP addresses can be introduced in the performance testing of network interconnect devices using IPv4 or IPv6 addresses, observing the limitations of the ranges of special purpose IPv4 and IPv6 addresses reserved for benchmarking measurements. Practical recommendations are given for the usage of pseudorandom source and destination IP addresses in the case of both IPv4 and IPv6 following the approach of RFC 4814 regarding the port numbers and also considering the effect of the growing number of ARP or NDP table entries.

Requirements Language The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

First, the potential ranges are examined in the case of IPv4 and IPv6, then considerations regarding the "optimal" number of IP addresses and the usage of gateways are made, and finally, a recommendation is given.

The 198.18.0.0/15 IPv4 address range was reserved for benchmarking tests. It is divided into two halves: 198.18.0.0/16 and 198.19.0.0/16 are to be used on the two sides of the test setup. Considering the requirement of regarding the IP addresses, the test suite SHOULD be first run with a single source and destination address pair. Then, the destination networks should be random using the 16-23 bits of the network addresses mentioned above. The Device Under Test (DUT) is assigned the first address of each network, and the Tester can be assigned, for example, the second address from each network. That is, 198.18.R.1/24 and 198.19.R.1/24 are assigned to the DUT; 198.18.R.2/24 and 198.19.R.2/24 are assigned to the Tester, where R is pseudorandom in the 0-255 interval. The above framework imposes serious limitations on the design of how multiple IP addresses can be used. It means that when, e.g., the very first networks (198.18.0.0/24 and 198.19.0.0/24) are used at each side of the test setup, the maximum range of the IP addresses assigned to the Tester can be 198.18.0.2/24-198.18.0.254/24 and 198.19.0.2/24-198.19.0.254/24, as shown in . When 256 destination networks are used, then the 16-23 bits identifying the destination networks also contribute to the entropy provided to the hash function. When only a single destination network is used, then the 16-23 bits can also be leveraged to generate a higher number of IP addresses, thus their ranges can be: 198.18.0.2/16-198.18.255.254/16 and 198.19.0.2/16-198.19.255.254/16 as shown in . In both cases, the Tester and the DUT are in the same networks, that is, they are connected directly without using gateways. Conversely, the corresponding network interfaces of the Tester and the DUT may be connected through gateways. Then the network interface of the DUT uses an IP address from a different network than the corresponding interface of the Tester, and the network interface of the DUT sends the packets to the gateway, which is set as the next hop router towards the network assigned to the corresponding interface of the Tester, as shown in . (It is noted that private IP addresses were used due to the insufficient IP address range reserved for benchmarking.)

| DUT: IPv4 router |-------------+ 198.18.0.1/24 | | 198.19.0.1/24 +----------------------------------+ ]]>

| DUT: IPv4 router |-------------+ 198.18.0.1/16 | | 198.19.0.1/16 +----------------------------------+ ]]>

| DUT: IPv4 router |-------------+ 10.0.0.1/24 | | 10.0.1.1/24 +----------------------------------+ routes in the DUT: destination network next hop ------------------- -------- 198.18.0.0/16 10.0.0.2 198.19.0.0/16 10.0.1.2 ]]>

The 2001:2::/48 IPv6 address range, which was reserved for benchmarking tests, is large enough. If it is split into two halves to be used on the two sides of the test setup as 2001:2::/49 and 2001:2:8000::/49, the ranges are abundant. Even if their first /56 subnets (2001:2::/56 and 2001:2:8000::/56) are enough to ensure 256 networks on each side of the test setup. As these networks are of /64 size, their host ID parts are vastly abundant. For convenience considerations, we recommend the usage of their 96-111 bits to generate potentially 65536 different IP addresses, as shown in in the case when a single destination network is used and the Tester and the DUT are connected directly, without gateways. (When needed, the 256 destination networks can be described by the 56-63, bits as mentioned before.) shows the case when gateways are used.

| DUT: IPv6 router |-------------+ / | | \ / +----------------------------------+ \ 2001:2::1/64 2001:2:0:8000::1/64 ]]>

| DUT: IPv6 router |-------------+ / | | \ / +----------------------------------+ \ 2001:2:0:1000::1/64 2001:2:0:9000::1/64 routes in the DUT: destination network next hop ------------------- -------- 2001:2::/64 2001:2:0:1000::2/64 2001:2:0:8000::/64 2001:2:0:9000::2/64 ]]>

On the one hand, the more IP addresses are used, the more entropy is ensured and thus the most even distribution of the load over the processing elements can be expected. However, on the other hand, the usage of multiple IP addresses has its costs when no gateways are used: multiple Address Resolution Protocol (ARP for IPv4) or Neighbor Discovery Protocol (NDP, for IPv6) table entries are used. Increasing them over a few thousand may have a deteriorating effect on the performance of the DUT. This effect does exist when gateways are used because the DUT sends the packets to the gateways, and thus only their IPv4 or IPv6 addresses needed to be stored in the ARP or NDP table of the DUT. It is noted that under typical operating conditions, a router is not connected directly to a high number of devices. If it is a backbone router, it is connected directly to several other routers. If it is a local router, it is connected directly to a single upstream router (or, at most, a few of them) and (through a switch) to the local hosts, the number of which is unlikely to be higher than a few thousand. In both cases, a high number of different IP addresses may provide entropy for hashing without causing pressure on the ARP or NDP tables of the router.

Based on the theoretical considerations in that it is desirable to use as high number of IP addresses as possible routers do not need to handle a high number of neighbors under typical operating conditions the usage of the maximum possible number of IP addresses and the test setups with gateways shown in and are recommended. It is noted that the maximum possible number of IPv4 addresses are 253 and 65533 when 8 bits and 16 bits are available. In the case of IPv6, it is 65536. This recommendation has been justified by the measurement results mentioned in .

The recommended solution has been implemented in siitperf as a proof of concept. Multiple IPv4 and IPv6 addresses are supported from commit number 165cb7f on September 6, 2023. The details of the implementation can be found in .

The effectiveness of the solution was also demonstrated in . OpenBSD was chosen as the operating system of the DUT. It uses the first type of RSS solution: only the IP addresses are used by the hash function. It was examined how much difference the usage of multiple IP addresses makes in the IPv4 and IPv6 throughput performance. It should be noted that IP packet forwarding under OpenBSD was single-threaded until version 7.1. The ChangeLog of OpenBSD 7.2 states, "Activated parallel IP forwarding, starting 4 softnet tasks but limiting the usage to the number of CPUs." The test setup for the IPv4 and IPv6 measurements was according to and , respectively. However, only 1000 different IP addresses were used at each side of the test setups to limit the potential performance degradation caused by the high number of ARP or NDP table entries. The DUT was a Dell PowerEdge R730 server with two 3.2GHz Intel Xeon E5-2667 v3 CPUs having 8 cores each, 8x16GB 2133MHz DDR4 SDRAM (accessed quad channel), and Intel X540-T2 10GbE network adapter. Hyper-threading was switched off in the BIOS. All tests were executed 10 times, and the median, minimum and maximum values of the throughput results were calculated. In the case of IPv4 packet forwarding, the usage of pseudorandom IP addresses caused a highly significant (more than 3-fold) performance increase compared to the case when fixed IP addresses were used. In the case of IPv6, the throughput values were significantly lower, and the increase caused by the usage of pseudorandom IP addresses was only about 50%, but it is still a well-visible difference. All the details of the measurements can be found in .

To examine how the number of IP addresses used affects the throughput, the same hardware described in was used, but the DUT had the Debian Linux 11.7 operating system to be able to fully utilize all CPU cores, and the first type of RSS was set using the "ethtool" command. The IPv4 and IPv6 test setups followed and , respectively. Two measurement series were performed: one with IPv4 and the other with IPv6. As for the IPv4 addresses, their number was doubled in the consecutive experiments: 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1k, 2k, 4k, 8k, 16k, 32k, and 64k-3 IPv4 addresses were used. Thus, the number of IPv4 address combinations (supplying entropy to RSS) were 1x1, 2x2, 4x4, 8x8, 16x16, 32x32, ... , 16kx16k, 32kx32k, and (64k-3)x(64k-3). The same number of IPv6 addresses were used, except that the last value was exactly 64k. The trend of the results followed the same pattern for IPv4 and IPv6. The throughput increased steeply during the first few consecutive experiments, then it was nearly constant for a relatively wide range of the number of IP addresses, and finally, it deteriorated significantly. Although the shallow observer could conclude that the above experienced nearly constant throughput for a wide range of the number of IP addresses, making it easy to choose a "good enough" number, the authors contend the opposite. On the one hand, in the general case, the number of CPU cores of the DUT may be significantly higher than in the above case. On the other hand, the beginning of the performance degradation may depend on several factors, such as ARP or NDP table implementation, the sizes of the given level (L1, L2, etc.) caches of the CPU, etc. Moreover, it is problematic if the testing method needs to be aware of the details of the internal structure and operation of the DUT. Finally, performing a series of measurements (like above) is highly time-consuming, and thus, it should be avoided.

To examine the throughput of IPv4 and IPv6 packet forwarding when gateways are used, the test setups shown in and were employed, respectively. And the test system was the same as in . The number of IPv4 and IPv6 addresses was 64k-2 and 64k, respectively. To have a basis for comparison, the IPv4 and IPv6 packet forwarding throughput measurements were also performed using single IPv4 and IPv6 address pairs plus pseudorandom port numbers with the second type of RSS setting, where port numbers are also included in the hash function. The two different tests gave approximately the same results in both the IPv4 and the IPv6 measurements.

Stateful technologies like stateful NAT44 or stateful NAT64 are out of the scope of this document. They are covered by .

The authors would like to thank Boris Hassanov for his review and comments.

This document does not make any request to IANA.

The same as in Section 13 of .

&RFC1918; &RFC2119; &RFC2544; &RFC4814; &RFC5180; &RFC8174; &RFC8219;

Initial version.

Measurement results added.

Minor update (one cited reference was published).

Test setups with gateways, measurement results, and final recommendations added. Grammar checking was done.

Addressed the review comments of Boris Hassanov. Recommendation for Testing with Multiple IP Addresses was moved from Section 6 to Section 2.4.