Userspace networking: beyond

the kernel bypass with RDMA!

Benoît Ganne, bganne@cisco.com

30/01/2020 1
 Why a native network driver?
• Why userspace networking? VPP IPv4 forwarding PDR, 1core, 2M routes
• Performance (avoid kernel overhead)
18
• Update network functions seamlessly (no reboot
required, containerization) 16 +23%
• Why your own network driver? 14
• Performance (metadata translation tax, feature 12
tax)

Mpps
• Ease-of-use (no reliance on hugepages, etc.) 10
8
• Why you should think twice?
• No integration with kernel (interface fully owned 6
by userspace) 4
• You care about rx/tx packets but device 2
initialization & setup is 95% of the work
• Hardware is hard (more on that later) 0
Native DPDK
Source: https://docs.fd.io/csit/master/report/vpp_performance_tests/packet_throughput_graphs/ip4-2n-skx-xxv710.html

30/01/2020 2
RDMA
• « Remote Direct Memory Access »
• Designed for message passing and HW RNIC
data transfer
• Has evolved to use Ethernet transport
(iWARP, RoCE)
Kernel RDMA uAPI
• Key properties

DMA
• Hardware offload
• Kernel bypass
libibverb
• Zero Copy data transfer
• High network bandwidth User

VPP
➔ Great for kernel networking!

30/01/2020 3
Extending RDMA for Ethernet Incoming packets are steered to Linux
netdev or userspace application based on
flows
• Not designed for efficient Ethernet
communication – but!
• Ethernet-capable HW (initially for HW RNIC
transport)
• High performance (200Gbps today)

DMA
• Kernel bypass with well established API RDMA
and native Linux kernel support Kernel
RDMA uAPI Netstack

DMA
• Why not extend it to support
userspace networking?
• Introduce new IBV_QPT_RAW_PACKET libibverb
queue pair type User
• Support for bifurcation with flow steering
• Keep your Linux netdev
• Support MACVLAN, IPVLAN model…
VPP

30/01/2020 4
 Using RDMA for Ethernet
How to send 20 Mpps with 1 CPU
1. Get a handle to the device you want to
use
2. Initialize queues
• Queue Pair (QP) = Submission Queue (SQ) +
Completion Queue (CQ)
• Protection Domain (PD) = where the NIC is
allowed to read/write data (packets)
3. Send packets
• Put Work Queue Elements (WQE – kind of
IOV) in SQ
• Notify new packets to send
• Poll CQ for completion
Full example at https://github.com/bganne/rdma-pktgen

30/01/2020 5
Going deeper with Direct Verbs
• RDMA user API is ibverb
• Simple enough, mostly standard,
open-source
• Not full performance (metadata
translation tax, feature tax)
• Direct Verbs
• ibverb extension to access DMA
ring-buffers directly
• Hardware-dependent!
• Setup done through ibverb, then
get DMA rings addresses

30/01/2020 6
VPP native RDMA driver
• ibverb version
• Available since 19.04
• ~ 20 Mpps L2-xconnect per core
• Direct Verb
• Development underway
• Hardware is hard: while trying to debug my driver I almost bricked my NIC
• Next
• Add support for hardware offloads (checksum offload, TSO)

30/01/2020 7
A call to action
• We love this model
• No need to write code boilerplate to initialize the NIC: we can focus on what matters
(rx/tx packets)
• Seamless integration with Linux kernel
• Great performance
• But is has limitations
• Need RDMA-capable NIC: must support Hardware security model, etc.
• Only supported on Mellanox for now
• Could other technologies enable this approach?
• Disclaimer: a bit outside of my domain knowledge here…
• vfio-mdev?
• AF_XDP?

30/01/2020 8