Scalability of IPv6 Transition Technologies for IPv4aaS

1.1. 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 BCP14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶

2.1. Measurement Method

[RFC8219] has defined a benchmarking methodology for IPv6 transition technologies. [I-D.lencse-bmwg-benchmarking-stateful] has amended it by addressing how to benchmark stateful NATxy gateways using pseudorandom port numbers recommended by [RFC4814]. It has defined a measurement procedure for maximum connection establishment rate and reused the classic measurement procedures like throuhput, latency, frame loss rate, etc. from [RFC8219]. We used two of them: maximum connection establishment rate and throughput to characterize the performance of the examined system.¶

The scalability of iptables is examined in two aspects:¶

• How its performance scales up with the number of CPU cores?¶
• How its performance degrades, when the number of concurrent sessions is increased?¶

              +--------------------------------------+
     10.0.0.2 |Initiator                    Responder| 198.19.0.2
+-------------|                Tester                |<------------+
| private IPv4|                         [state table]| public IPv4 |
|             +--------------------------------------+             |
|                                                                  |
|             +--------------------------------------+             |
|    10.0.0.1 |                 DUT:                 | 198.19.0.1  |
+------------>|        Sateful NATxy gateway         |-------------+
  private IPv4|     [connection tracking table]      | public IPv4
              +--------------------------------------+

The test setup in Figure 1 was followed. The two devices, the Tester and the DUT (Device Under Test), were both Dell PowerEdge R430 servers having two 2.1GHz Intel Xeon E5-2683 v4 CPUs, 384GB 2400MHz DDR4 RAM and Intel 10G dual port X540 network adapters. The NICs of the servers were interconnected by direct cables, and the CPU clock frequecy was set to fixed 2.1 GHz on both servers. They had Debian 9.13 Linux operating system with 4.9.0-16-amd64 kernel. The measurements were performed by siitperf [LEN2021] using the "stateful" branch (latest commit Aug. 16, 2021). The DPDK version was 16.11.11-1+deb9u2. The version of iptables was 1.6.0.¶

The ratio of number of connections in the connection tracking table and the value of the hashsize parameter of iptables significantly influences its performance. Although the default setting is hashsize=nf_conntrack_max/8, we have usually set hashsize=nf_conntrack_max to increase the performance of iptables, which was crucial, when high number of connections were used, because then the execution time of the tests was dominated by the preliminary phase, when several hundereds of millions connections had to be established. (In some cases, we had to use different settings due to memory limitations. The tables presenting the results always contain these parameters.)¶

The size of the port number pool is an important parameter of the bechmarking method for stateful NATxy gateways, thus it is also given for all tests.¶

2.2. Performance scale up against the number of CPU cores

To examine how the performance of iptables scales up with the number of CPU cores, the number of active CPU cores was set to 1, 2, 4, 8, 16 using the "maxcpus=" kernel parameter.¶

The number of connections was always 4,000,000 using 4,000 different source port numbers and 1,000 different destination port numbers. Both the connection tracking table size and the hash table size was set to 2^23.¶

The error of the binary search was chosen to be lower than 0.1% of the expected results. The experiments were executed 10 times.¶

Besides the connection establishment rate and the throughput of iptables, also the throuhput of the IPv4 packet forwarding of the Linux kernel was measured to provide a basis for comparison.¶

The results are presented in Figure 2. The unit for the maximum connection establishment rate is 1,000 connections per second. The unit for throughput is 1,000 packets per second (measured with bidirectional traffic, and the number of all packets per second is displayed).¶

num. CPU cores          1          2          4          8         16
src ports           4,000      4,000      4,000      4,000      4,000
dst ports           1,000      1,000      1,000      1,000      1,000
num. conn.      4,000,000  4,000,000  4,000,000  4,000,000  4,000,000
conntrack t. s.      2^23       2^23       2^23       2^23       2^23
hash table size      2^23       2^23       2^23       2^23       2^23
c.t.s/num.conn.     2.097      2.097      2.097      2.097      2.097
num. experiments       10         10         10         10         10
error                 100        100        100      1,000      1,000
cps median          223.5      371.1      708.7      1,341      2,383
cps min             221.6      367.7      701.7      1,325      2,304
cps max             226.7      375.9      723.6      1,376      2,417
cps rel. scale up       1      0.830      0.793      0.750      0.666
throughput median   414.9      742.3      1,379      2,336      4,557
throughput min      413.9      740.6      1,373      2,311      4,436
throughput max      416.1      746.9      1,395      2,361      4,627
tp. rel. scale up       1      0.895      0.831      0.704      0.686
IPv4 packet forwarding (using the same port number ranges)
error                 200        500      1,000      1,000      1,000
throughput median   910.9      1,523      3,016      5,920     11,561
throughput min      874.8      1,485      2,951      5,811     10,998
throughput max      914.3      1,534      3,037      5,940     11,627
tp. rel. scale up       1      0.836      0.828      0.812      0.793
throughput ratio (%) 45.5       48.8       45.7       39.5       39.4

Whereas the throughput of IPv4 packet forwarding scaled up from 0.91Mpps to 11.56Mpps showing a relative scale up of 0.793, the throuhput of iptables scaled up from 414.9kpps to 4,557kpps showing a relative scale up of 0.686 (and the relative scale up of the maximum connection establishment rate is only 0.666). On the one hand, this is the price of the stateful operation. On the other hand, this result is quite good compared to the scale-up results of NSD (a high performance authoritative DNS server) presented in Table 9 of [LEN2020], which is only 0.52. (1,454,661/177,432=8.2-fold performance using 16 cores.) And DNS is not a stateful technology.¶

2.3. Performance degradation caused by the number of sessions

To examine how the performance of iptables degrades with the number connections in the connection tracking table, the number of connections was increased fourfold by doubling the size of both the source port number range and the destination port number range. Both the connection tracking table size and the hash table size was also increased four fold. However, we reached the limits of the hardware at 400,000,000 connections: we could not set the size of the hash table to 2^29 but only to 2^28. The same value was used at 800,000,000 connections too, when the number of connections was only doubled, because 1.6 billion connections would not fit into the memory.¶

The error of the binary search was chosen to be lower than 0.1% of the expected results. The experiments were executed 10 times (except for the very long lasting measurements with 800,000,000 connections).¶

The results are presented in Figure 3. The unit for the maximum connection establishment rate is 1,000,000 connections per second. The unit for throughput is 1,000,000 packets per second (measured with bidirectional traffic, and the number of all packets per second is displayed).¶

num. conn.       1.56M   6.25M     25M     100M     400M     800M
src ports        2,500   5,000  10,000   20,000   40,000   40,000
dst ports          625   1,250   2,500    5,000   10,000   20,000
conntrack t. s.   2^21    2^23    2^25     2^27     2^29     2^30
hash table size   2^21    2^23    2^25     2^27     2^28     2^28
num. exp.           10      10      10       10       10        5
error            1,000   1,000   1,000    1,000    1,000    1,000
n.c./h.t.s.      0.745   0.745   0.745    0.745    1.490    2.980
cps median       2.406   2.279   2.278    2.237    2.013    1.405
cps min          2.358   2.226   2.226    2.124    1.983    1.390
cps max          2.505   2.315   2.317    2.290    2.050    1.440
thorughput med.  5.326   4.369   4.510    4.516    4.244    3.689
thorughput min   5.217   4.240   3.994    4.373    4.217    3.670
thorughput max   5.533   4.408   4.572    4.537    4.342    3.709

The performance of iptables shows degradation at 6.25M connections compared to 1.56M connections very likely due to the exhaustion of the L3 cache of the CPU of the DUT. Then the performance of iptables is fearly constant up to 100M connections. A small performance decrease can be observed at 400M connections due to the lower hash table size. A more significant performance decrease can be observed at 800M connections. It is caused by two factors:¶

• on average, about 3 connections were hashed to the same place¶
• non NUMA local memory was also used.¶

We note that the CPU has 2 NUMA nodes, cores 0, 2, ... 14 belong to NUMA node 0, and cores 1, 3, ... 15 belong to NUMA node 1. The maximum memory consumption with 400,000,000 connections was below 150GB, thus it could be stored in NUMA local memory.¶

Therefore, we have pointed out important limitations of the stateful NAT44 technology:¶

• there is a performance decrease, when approaching hardware limits¶
• there is a hardware limit, beyond which the system cannot handle the connections at all (e.g. 1600M connections would not fit into the memory).¶

Therefore, we can conclude that, on the one hand, a well tailored hashing may guarantee an excellent scale-up of stateful NAT44 regarding the number of connections in a wide range, however, on the other hand, stateful operation has its limits resulting both in performance decrease, when approaching hardware limits and also in inability to handle more sessions, when reaching the memory limits.¶

A.1. 00

Initial version: scale up of iptables.¶