The most direct role of the DPU is to serve as an offload engine for the CPU , taking over infrastructure layer services such as network virtualization and hardware resource pooling, and releasing the computing power of the CPU to upper-layer applications. Taking network protocol processing as an example, about 4 Xeon CPU cores are required to process a 10G network at a wire speed. That is to say, only half of the computing power of an 8 -core high-end CPU can be used for network packet processing . If 40G and 100G high-speed networks are considered, the performance overhead is even more unbearable. Amazon calls these expenses " Datacenter Tax " , that is, the computing resources that will be occupied by accessing network data before running business programs. The AWS Nitro product family aims to offload data center overhead ( providing remote resources for virtual machines, encryption and decryption, fault tracking, security policies and other service programs) from the CPU to the Nitro accelerator card, which will release 30% of the original resources for upper-layer applications. The computing power used to pay " Tax " !
DPU can become a new data gateway, raising security and privacy to a new level. In the network environment, the network interface is the ideal boundary of privacy, but the encryption and decryption algorithms are expensive, such as the national secret standard asymmetric encryption algorithm SM2 , hash algorithm SM3 and symmetric block cipher algorithm SM4 . If the CPU is used for processing, only a small amount of data can be encrypted. In the future, as the business carried by the blockchain gradually matures, running the consensus algorithm POW and verifying signatures will also consume a lot of CPU computing power. And all of these can be achieved by solidifying it in the DPU , and even the DPU will become a root of trust.
DPU can also be the entrance of storage, localizing distributed storage and remote access. As the performance-to-performance ratio of SSDs becomes more and more acceptable, it is possible to migrate part of the storage to SSD devices. The traditional SATA protocol for mechanical hard disks is not suitable for SSD storage. Therefore, connecting SSDs to the system through local PCI e or high-speed network becomes Mandatory technical route. NVMe ( Non Volatile Memory Express ) is a high-speed interface standard protocol for accessing SSD storage. PCIe can be used as the underlying transmission protocol to fully utilize the bandwidth advantages of SSDs . At the same time, in a distributed system, it can also be extended to InfiniBand , Ethernet , or Fiber channel nodes through the NVMe over Fabrics ( NVMe-o F ) protocol, and RDMAShared storage and remote access in the form of storage. These new protocol processing can be integrated in the DPU to achieve transparent processing to the CPU . Furthermore, the DPU may take on the role of various interconnection protocol controllers to achieve a better balance point in terms of flexibility and performance.
The DPU will become a sandbox for algorithm acceleration and the most flexible accelerator carrier. DPU is not completely a solidified ASIC . Under the pavement of data consistency access protocols such as CPU , GPU and DPU advocated by standard organizations such as C XL and CCIX , it will further clear the programming barriers of DPU and combine FPGA and other programmable devices , Customizable hardware will have more room to play, " software and hardware " will become the norm, and the potential of heterogeneous computing will be fully realized due to the popularity of various DPUs . Corresponding DPUs may appear in the fields where " Killer Application " appears , such as traditional database applications such as OLAP , OLTP , 5G edge computing, intelligent driving V2X , etc.
Source: White Paper on Dedicated Data Processor (DPU), Institute of Computing Technology, Chinese Academy of Sciences, Yan Guihai, etc.
Full text download: https://www.yusur.tech/zkls/zkys/dpu_whitepaper.html