Capabilities

Link-Inline™ Software Features

Platform
Napatech SmartNIC
Content Type
Feature Description
Capture Software Version
Link-Inline™ Software 3.2

Napatech Link-Inline™ Software ensures that all programmed flows are processed on the SmartNIC for maximizing overall system performance.

Lookup table capacity

The lookup table for flows and policies is stored in onboard SDRAM. 140 million records can be stored on a NT200A02 SmartNIC and a NT400D11 SmartNIC (64-byte records, using 12-Gbyte onboard SDRAM).
Note: Frames from the receive pipeline are not stored in onboard SDRAM before transferred to the host memory or transmitted on the TX port.

Lookup performance

Flow lookups are stateless as flow records are not updated whereas policy lookups are stateful as updating the leaky-bucket state and usage counters for metering are carried out. Stateless lookups (flow lookups) guarantee a rate of 130 million lookups per second, while stateful lookups (policy lookups) guarantee a rate of 85 million lookups per second.

These lookup rates are calculated under worst case scenarios, assuming no cache hits and a fully filled lookup table. However, in practical scenarios, lookup rates tend to be much higher. This is because the SmartNIC caches recently used flows and policies, and the flow table would not reach full capacity as demonstrated in the following figures.


Graphs showing test results of NT400D11 lookup rates with stateless lookups and stateful lookups.


Graphs showing test results of NT200A02 lookup rates with stateless lookups and stateful lookups.

These figures illustrate the performance of flow lookup operations. The test was conducted under worst case scenarios using a round-robin traffic pattern. This involves frames with a sequence of different flows.

where:
  • Blue plot: Lookup rates when 1-flow lookup per frame is performed and no additional policy lookups are performed.
  • Red plot: Lookup rates when 1-flow/1-policy lookups per frame are performed, and 1000 policies are registered in the lookup table.
  • Green plot: Lookup rates when 1-flow/2-policy lookups per frame are performed, and 1000 policies are registered in the lookup table.
  • Turquoise plot: Lookup rates when 1-flow/1-policy lookups per frame are performed, and 1 million policies are registered in the lookup table.
  • Purple plot: Lookup rates when 1-flow/-2 policy lookups per frame are performed, and 1 million policies are registered in the lookup table.

The blue plot displays the lookup rates without any policy lookups (stateless lookups), which serves as a baseline reference. For the NT400D11 SmartNIC, upon populating the lookup table with approximately 10 million flows, the lookup rate exceeds 250 million lookups per second. Upon further scaling the lookup table with approximately 100 million flows, the lookup rate shows approximately 150 million lookups per second. For the NT200A02 SmartNIC, upon populating the lookup table with approximately 10 million flows, the observed lookup rate shows above 180 million lookups per second. With 100 million flows in the lookup table, the lookup rate shows within the range of 140 to 150 million lookups per second.

Naturally, the lookup rates decrease as more flows and policies are added to the lookup table. When the SmartNIC caches flows and policies, the lookup rates for the NT400D11 SmartNIC reach around 300 million frames per second or 200 million frames per second, while the lookup rates for the NT200A02 SmartNIC reach around 220 million frames per second, 150 million frames per second or 100 million frames per second. Once the cache memory is exhausted, the lookup rates decrease.

The green plot indicates the lookup rates a maximum of 200 million frames per second for the NT400D11 SmartNIC and 100 million frames per second for the NT200A02 SmartNIC. This limitation is due to reaching the maximum processing capacity of the FPGA pipeline in the SmartNIC, which handles 3 lookups per frame (1-flow lookup and 2-policy lookups). The FPGA pipeline of the NT400D11 SmartNIC can process a maximum of 600 million lookups per second, while the NT200A02 can handle a maximum of 300 million lookups per second.

The NT400D11 SmartNIC demonstrates higher lookup rates due to the higher clock frequency for its FPGA pipeline compared to that of the NT200A02 SmartNIC.

Fast path forward latency

The following figure shows the test results for fast path forward latency also known as port-to-port latency.


Graphs showing test results of NT400D11 and NT200A02 latency.

The figure illustrates latency under varying traffic loads, scaling up to 200 Gbits/s. The blue plot represents latency for the NT400D11 SmartNIC, showing an average latency range of 3.1 to 3.5 µs. The red plot shows latency for the NT200A02 SmartNIC, averaging between 4.6 to 4.9 µs.

During the test, the SmartNIC efficiently manages traffic with offload functionalities, including frame classification, flow lookups, decapsulation/encapsulation and frame modification for each frame. In addition, the test was conducted under worst-case scenarios using a round-robin traffic pattern with frames sequenced across multiple flows.

Other capabilities

  • Learning rate: Higher than 3 million flows per second.
  • Unlearning rate: Higher than 3.5 million flows per second.