What is SmartNIC?
SmartNICs are the latest generation of LAN Cards which has high throughput in the range of 40-200gbps and packs a lot of feature in packet processing,
As compared to Traditional NICS SmartNICs allow functionality like IO Virtualization (SRIOV), PCIE-passthrough. They are Cloud &SDN compatible i.e Openstack, VMware, Azure, CORD, ONOS
With ever increasing demand for higher speed (pkt throughput), CPU cycles consumption for networking is on rise. This leaves lesser CPU cycles for General purpose computing.
Hence the concept of SmartNIC’s which can take up CPU-intensive packet processing tasks. E.g.
A Lookup operation especially having 100k plus entries is a costly operation on CPU.
CPU lookups are software-based lookups which are much slower as compared to ASIC or FPGA.
An ASIC/FPGA can execute same operation 10X faster.
Many such Algorithms/ pkts can be process in different LUTs which are driven by same clock.
ASIC/FPGA can do crypto operations 100X faster. E.g MD5, HMAC, SHA, AES.
Multiple Parallel Circuits can execute data driven by Clock.
Another Possibility is: Single pkt can be operated at L1-L4 levels in a highly parallel manner.
Interrupt Driven Mechanism chokes the CPU cores. Major effort is lost in shuffling the tasks/process.
Linux networking stack takes an avg about 1200 CPU cycles= 600ns for end to end pkt processing.
The interrupt-based mechanism of NIC card hasn’t changed since a very long time in computing history.
Also the CPU cycles involved in doing L1-L4 packet processing are Huge
SMART-NOMICS in cloud computing:
Public Cloud vendors like
aws, GCP, Azure need to meet quality networking infrastructure std
required for hosting enterprise grade software in Cloud.
As the main business of Cloud Providers is to sell VM’s(vcpu), high CPU consumption for networking eats into their revenue.
As per Data released by Amazon-aws a 40Gbps throughput on 32 core system consume 18% of CPU cycles,
18% of 32 cores = 6 cores.
On an average 1core vcpu VM generates $1000 revenue/year.
So 6 cores needs to be reserved for Networking throughput, which takes away $6000 revenue per server box.
So a modest 1000 rack cloud will incur a revenue loss of $6million dollars annually.
Cost advantage in ASIC/FPGA.
ASIC based NICs are the defacto std, they are designed to.
1.Do a very specific job at very good rate.
2.Power-Consumption is low @Fast Speed.
3.Cheap to manufacture.
Now a days improved version of SMARTNIC’s are designed in combination with processors.
Xilinx SmartNIC come with UltraScale series fpga and Zynq processors.
Intel N3000 series has arm-based processors.
Mellanox Blue-field has up to 8 cores ARM processors.
Feature Highlights of SmartNIC.
|Datapath Offload||MPLS, VLAN, NVGRE, OFFLOAD|
|Network Acceleration and Offload||TCP/UDP IP stateless offload, QoS Enhancement, RAN HARQ offload|
|Image-Processing offload||Image compression/decompression before transmission.|
To read more please refer to this link:
Do Let us know your views in comment section.I will Post next series on.
What Openstack supports in SmartNIC Openstack Cyborg.AND P4 The next frontier in democratizing ASIC Programming.