This example gives a skeleton ECMP implementation, using an action profile. A P4 action profile is mapped to an indirect table in bmv2. More precisely, in our case, the P4 objects ecmp_group and ecmp_action_profile are mapped to one bmv2 indirect match table. This has the limitation that sharing action profiles between P4 tables is not supported in bmv2.

Running the demo

As always, you can start the switch with ./run_switch.sh (and wait until READY is displayed). We program the dataplane with the following CLI commands (from [commands.txt] (commands.txt)):

01. table_indirect_create_group ecmp_group
02. table_indirect_create_member ecmp_group _nop
03. table_indirect_create_member ecmp_group set_ecmp_nexthop 1
04. table_indirect_create_member ecmp_group set_ecmp_nexthop 2
05. table_indirect_create_member ecmp_group set_ecmp_nexthop 3
06. table_indirect_add_member_to_group ecmp_group 1 0
07. table_indirect_add_member_to_group ecmp_group 3 0
08. table_indirect_set_default ecmp_group 0
09. table_indirect_add_with_group ecmp_group 10.0.0.1 => 0
10. table_dump ecmp_group

These commands do the following:

create a group, which receives group handle 0
create a member with action _nop and no action data (member handle 0)
create a member with action set_ecmp_nexthop and action data port=1 (member handle 1)
create a member with action set_ecmp_nexthop and action data port=2 (member handle 2)
create a member with action set_ecmp_nexthop and action data port=3 (member handle 3)
add member 1 to group 0
add member 3 to group 0
set member 0 (_nop) has the default for table ecmp_group
add an entry to table ecmp_group, which maps destination IPv4 address 10.0.0.1 to group 0 (which includes members 1 and 3)
simply dump the table

These commands are sent automatically when you run ./run_switch.sh. The last command should display the following information:

ecmp_group:
0: 0a000001 => group(0)
members:
0: _nop - 
1: set_ecmp_nexthop - 1,
2: set_ecmp_nexthop - 2,
3: set_ecmp_nexthop - 3,
4: set_ecmp_nexthop - 4,
groups:
0: { 1, 3, }

When we send a packet to the switch (on any port) with destination address 10.0.0.1, the egress port will be set to either 1 or 3. This is because we only added members 1 and 3 to the group. Which port / member is used for a given packet is determined by computing a hash over the IP source address and the IP protocol number. See the P4 program for more details.

After starting the switch, run sudo python send_and_count.py. This Python script will send 200 packets to the switch and count the pakcets coming out of ports 2 and 3. The two numbers should be almost the same:

Sending 200 packets on port 0
port 1: 106 packets (53%)
port 2: 0 packets (0%)
port 3: 94 packets (47%)
port 4: 0 packets (0%)