maglev

Maglev Load Balancer

Building

mkdir build_release
cd build_release
cmake -DCMAKE_BUILD_TYPE=Release -D CMAKE_C_COMPILER=gcc-9 -D CMAKE_CXX_COMPILER=g++-9 ..
make

Running

Ensō

To run the Ensō version of Maglev, run the following command from within the build_release directory:

sudo ./bin/enso_maglev -l 0-<nb_cores-1> -- <nb_cores> <queues_per_core> <nb_backends>

For example, to run with 2 cores, 4 queues per core and 1000 backends, run:

sudo ./bin/enso_maglev -l 0-1 -- 2 4 1000

Also note that Maglev uses hash of the 5-tuple to direct packets to cores. Remember to set the number of fallback queues on the NIC depending on the number of cores and queues per core. For example, if you are running with 2 cores and 4 queues per core, set the number of fallback queues to 8 using the enso command:

enso <enso path> --host <host> --fpga <fpga id> --fallback-queues 8

You may also set the number of fallback queues using the JTAG console. Refer to the docs for more information.

DPDK

To run the DPDK version of Maglev, first do the things required by DPDK, such as setup hugepages and bind the NIC to DPDK. Then run the following command from within the build_release directory:

sudo ./bin/dpdk_maglev -l 0-<nb_cores-1> -m 4 -a <NIC pcie address> -- --q-per-core <queues_per_core> --nb-backends <nb_backends>

For example, to run with 2 cores, 4 queues per core, 1000 backends with the NIC at 0000:01:00.0, run:

sudo ./bin/dpdk_maglev -l 0-1 -m 4 -a 0000:01:00.0 -- --q-per-core 4 --nb-backends 1000