Wire-Speed Hardware-Assisted Traffic Filtering with Mainstream Network Adapters - Advances in Network-Embedded Management and Applications - page 80

queue id> . Currently all configured filters must have the same mask
defined in 82599.

The 82599 adapter is quite different from many FPGA-based NICs as it does
not use a TCAM (Ternary Content Addressable Memory) for handling filters.
This means that a filter is configured by setting up specific NIC registers and,
therefore, that the last configured filter overwrites the previous register value.
For this reason, it is not possible to read from the NIC all configured filters,
and therefore the driver has to maintain the list of configured filters. The
advantage of this approach is that, contrary to many FPGA-based NICs where
setting a filter requires card reconfiguration, in 82599 setting a filter is
extremely fast and from the application point of view it takes as long as the
setsockopt() system call necessary to pass the filter specification to the
kernel, making this NIC usable in environments where filter configuration has
to be dynamically changed.

5 Use Cases and Validation

Validation has been performed using an IXIA XM12 10 Gbit traffic generator
and a NUMA computer using a single 6-core Xeon® X5650 (Westmere) CPU
at 2.67GHz. In all tests we have injected IPv4 UDP traffic with random
payload at wire speed, and compared the number of packets sent by the traffic
generator with those reported by pfcount, a simple packet-counting
application running on top of PF_RING. pfcount spawns and binds a thread
per core (i.e. thread X is bound to core X). The injected traffic contained 6
flows, each balanced to an individual core using hardware filtering rules.
Packets have been captured using the standard NAPI-based 82599 driver
enhanced with PF_RING and hardware filtering support.

Table 1. Hardware vs. Software Filtering Comparison

Frame Size
(Bytes)

Test 1 Test 2

Software
Filter (Captur
e Rate)

Hardware
Filter
(Capture Rate)

Software
Filter (CPU
Load)

Hardware
Filter
(CPU Load)

64 5.7% 6.3% 95.6% None

128 10.0% 11.6% 95.4% None

256 19.5% 23.2% 98.7% None

512 37.4% 42.3% 3.5% None

1024 99.8% 100% 3.3% None

1518 99.6% 100% < 0.1% None