Look at the top 10 capabilities of alicloud cipu

Catalog

CIPU

With the traditional SmartNIC Manufacturer ratio , The cloud computing giant is entirely to meet its own cloud computing needs . In their view ,SmartNIC Is a complete calculation 、 Storage 、 Network integrated processing unit , It just depends on how you want to unload the specific business , The specific business is best known only by the cloud vendors themselves .

therefore , For cloud vendors , Since the research SmartNIC It seems to be the right way , Its “ Smart network card ” The concept of has been very weak , It's not just a NIC So simple , It's deep in their blood .

2017 year , Alibaba cloud has started to deploy completely self-developed X-Dragon MOC card .

stay 2022 At the Alibaba cloud summit in , Released a special processor for Cloud Data Center CIPU（Cloud infrastructure Processing Units, Cloud infrastructure processors ）. hold IDC Calculation 、 Storage 、 Dedicated business processors with cloud based network infrastructure and hardware acceleration .

CIPU The positioning of

CIPU In general scalar computation and AI Vector computing and other business areas , To complete XEON ALU Calculative sum GPU stream processor Of offloading Obviously not .CIPU Common business characteristics of the appropriate business type ： In data flow （ Move ） In the process , Through deep vertical hardware and software co design , Minimize data movement , So as to improve the computing efficiency .

therefore ,CIPU The main work from the perspective of computer architecture is ： Optimize data hierarchy between and within cloud computing servers cache、 Memory and storage access efficiency .

CIPU Random heterogeneous computing （ On the road to networking and storage ） = Near network computing （in networking computing）+ Near memory computing （in storage computing）

CIPU Of 10 Great ability

IO Hardware device virtualization

adopt VT-d Front support technology , Achieve high performance IO Hardware device virtualization .

Also consider the public cloud OS Ecological compatibility , The device model should be as compatible as possible . So the implementation is based on virtio-net、virtio-blk、NVMe And other industry standards IO Device model , Become a must .

At the same time, notice IO High performance of equipment , So in PCIe Optimization at the protocol level is crucial . How to reduce PCIe TLP Traffic volume 、 Reduce Guest OS Number of interrupts （ At the same time, the delay demand is balanced ）, Realize flexible hardware queue resource pooling , new IO Business programmability and configurable flexibility , It's a decision IO The key to the realization of hardware virtualization .

VPC overlay network Hardware acceleration

The business requirements of network virtualization are ：

• demand 1： Bandwidth and line speed processing capacity ;
• demand 2： The acme E2E Low delay and low delay jitter ;
• demand 3： High without packet loss pps Forwarding capability .

The current challenges of traditional solutions ：

1. 100Gbps+ Large bandwidth data mobility , Lead to “ Von Neumann memory wall ” Outstanding problems ;
2. CPU Scalar processing network virtualization services , The parallelism bottleneck is obvious ;
3. Software based data path processing , Delay jitter is difficult to overcome .

here , The business demand based on hardware forwarding acceleration was born , The technical realization level can be divided into ：

• Be similar to MNLX ASAP、Intel FXP、Broadcom trueflow Based on configurable ASIC Forwarding technology ： It has the highest performance to Watt ratio and the lowest forwarding delay , But the business flexibility is limited ;
• be based on many core Of NPU technology ： It has certain forwarding business flexibility , however PPA（power-performance-area） Efficiency and forwarding delay cannot be configured ASIC competition .
• FPGA Reconfigurable logic implements forwarding technology ：time to market Ability has great advantages , But for 400Gbps/800Gbps Forwarding service , It's a big challenge .

EBS Distributed storage access hardware acceleration

Public cloud storage should be realized 9 individual 9 Data persistence for , And the computing and storage should meet the elastic business requirements , It will inevitably lead to the separation of deposit and settlement .EBS（ Alibaba cloud block storage ） Must be on the nose Initiator High performance 、 Low delay access to the tail of the distributed storage Target.

Specific demand level ：

1. EBS As real-time storage , Must be realized E2E Extremely low latency and extremely P9999 Delay jitter ;
2. Realize line speed storage IO forward , Such as 200Gbps Under the network environment 6M IOPS;
3. new generation NVMe Hardware IO virtualization , While meeting the needs of shared disk business , solve PV NVMe Semi virtualization IO Performance bottleneck .

Local storage virtualization hardware acceleration

The local store , Although there is no such thing as EBS 9 individual 9 Data persistence and reliability , But at low cost 、 High performance 、 It still has advantages in terms of low delay , For calculation Cache、 Big data and other business scenarios are just needed .

If the local disk is virtualized , bandwidth 、IOPS、 Zero attenuation of time delay , At the same time, it has one empty and many 、QoS Isolation 、 O & M capability , It is the core competitiveness of local storage virtualization hardware acceleration .

elastic RDMA

RDMA The network HPC、AI、 big data 、 database 、 Storage, etc Data centric In business , Playing an increasingly important technical role . so to speak ,RDMA The Internet has become Data centric The key to business differentiation .

Specific needs ：

• Cloud based Overlay Large scale network deployment ,Overlay Where the network can reach ,RDMA The network can reach ;
• RDMA verbs ecology 100% compatible ,IaaS Zero code modification is the key to business success ;
• Very large scale deployment , Tradition RoCE Technology based on PFC etc. Data center bridging technology , In many aspects, such as network scale and switching network operation and maintenance , It's unsustainable . Cloud elasticity RDMA Technology needs to get rid of PFC And lossless network dependency .

Security hardware acceleration

Continue to strengthen hardware trusted technology 、VPC The east-west traffic is fully encrypted 、EBS And local disk virtualization data full encryption , Hardware based Enclave Technology, etc. , It is the key for cloud manufacturers to continuously improve cloud business competitiveness .

Cloud O & M capability support

The core of cloud computing is Service（ As a service ）, So as to realize the user's understanding of IT Free operation and maintenance of resources . and IaaS The core of the O & M capability of elastic computing is ：

• Lossless hot upgrade capability of all business components
• Lossless hot migration capabilities of virtual machines .

Elastic bare metal support

Elastic bare metal must realize the following eight key business characteristics at the specific definition level . meanwhile , Cloud computing elastic business inevitably requires elastic bare metal 、 virtual machine 、 Pooling production and scheduling of computing resources such as security containers .

CIPU Pooling capacity

Considering the general calculation and AI Computing on the network 、 The demand for storage and computing power varies greatly ,CIPU Must have pooling capability . General calculation passed CIPU Pool technology , Marked improvement CIPU Resource utilization , So as to enhance the core competitiveness at the cost level ; At the same time, it can CIPU Under the technical architecture system , Satisfy AI And other high bandwidth services .

Computing virtualization support

The business features of computing virtualization and memory virtualization are enhanced , Cloud vendors are all interested in CIPU There are many core requirements defined .