arithmetic power can no longer keep up with the growth of bandwidth, and the arithmetic power originally used to process business is heavily wasted on processing network data and infrastructure business (OVS, NFV), and the performance by way of CPU software emulation can no longer meet the demand, and server performance has reached a bottleneck, and many customers in the market may consider two options:
On traditional smart NICs, the processing of the first packet is on the CPU, and the flow table is sent down to the NIC chip still occupies CPU resources, and the interface and related protocols need to be defined for the offloaded service, which is complicated to implement and difficult to locate the problem.
On the other hand, on the SmartNIC of DPU architecture, full offload can be realized without much business interface between each other, and the positioning problem is clear. It is equivalent to having dedicated computing resources to deal with the control plane associated with the smart NIC, which can realize the full offload of the control plane + forwarding plane, and the processing performance of DPU is at an absolute advantage compared with the NIC of FPGA or SoC architecture.
Take cloud gateway access as an example. Under the premise of ensuring the total number of access users, the construction solution of introducing DPU architecture intelligent NIC has obvious advantages in rack space occupation, total power consumption and cost compared with the construction solution of pure 2U server. Among them, the rack space saving is more than 2/3, the total power consumption saving is more than 1/2, the construction cost saving is 8W, while the average number of users per U access on the rack is more than 3 times.
Asterfusion Helium SmartNIC based on Marvell OCTEON TX CN9670 which contains 24 -core ARM processors with 2 ×100G QSFP28 ports/ 4×25G SFP28 ports,along with multiple integrated hardware-acceleration co-processors that accelerates workloads for data center, cloud, Web2.0, security and storage scenarios.
The eBPF function is offloaded to the smart NIC. The ARM processor carried on the smart NIC implements a separate backend, compiles the intermediate eBPF bytecode into the instruction code of the ARM processor architecture, loads it into the NIC RAM (the smart NIC contains DRAM, so MAP can also be offloaded), and the smart NIC loads the eBPF program from the NIC RAM and runs it
The eBPF processes the data packets and returns the metadata and statistics to the host application.
For more information, visit: https://asterfusion.com/product/helium_dpu/ or send email to sales@asterfusion.com