The difference between ordinary switch and data center switch
Ordinary core switch is mainly aimed at meeting the interconnection and interoperability, and it is impossible to achieve accurate identification and control of services. It is impossible to respond quickly and lose packets in large business situations, and the continuity of services cannot be guaranteed; the reliability of the system mainly depends on the reliability of the equipment.
data center-level switches are characterized by high-quality service guarantee and control and identification capabilities, and end-to-end flow control and backpressure mechanisms to ensure the stability and reliability of data transmission and to smooth out network surges. Higher reliability and security, simpler networking methods, and faster business deployment.
The network surge capacity and service scheduling of the data center
To solve the key performance problems such as high-density application scheduling of data centers and traffic surge burst buffering, it is inevitable to carry out technological innovation in the infrastructure design of the switching platform.
first provides hardware-based traffic management capabilities on the switching platform. Large-capacity cache matches dense hardware scheduling queues, expanding scheduling capabilities to tens of thousands of queues. Once the upper-level application data flow enters the corresponding hardware queue, a large-scale (far more than 8 queues) data center-level service scheduling capabilities can be achieved.
Another technological change is to change the exit port cache method of traditional switching systems, and adopt a distributed ingress cache architecture. In traditional outgoing port caching, the service burst capacity of the entire system is determined only by the cache size that can be allocated by the outgoing port, so the capacity is fixed. The traffic reaches a certain burst limit, that is, the instantaneous burst data volume exceeds the outgoing port cache size, and the entire system begins to lose packets of and .
distributed caching technology adopts an architecture that is different from the traditional method. During normal forwarding, the outgoing port forwards data at a line speed of 10 Gigabits. When multiple burst traffic from 10 Gigabits to 10 Gigabits starts to exceed 10 Gigabits, the number of message mark queues on the outgoing port will grow rapidly, reaching a certain threshold (entering a quasi-congested state), end-to-end (E2E) flow control information is generated to each traffic source (ingress) port. The ingress port cache starts to cache burst traffic locally and stops sending data to the exit, while the exit still sends instantaneous over 10 Gigabit burst traffic at a line speed of 10 Gigabits. When the exit port releases the quasi-congestion state, the E2E flow control notification ingress cache will forward the retained data normally.
The entire distributed cache mechanism is coordinated by the traffic manager at the hardware level, without software participation, and therefore works at the system clock level. The size of each ingress cache requires a burst traffic cache capacity of 200 milliseconds under 10 Gigabit full-line speed. Therefore, when bursts cause instantaneous congestion, the cache capacity of N ports forwarding to one port is N*200 milliseconds, which is essentially improved compared with the traditional outgoing port cache fixed capability. Moreover, after testing, the actual performance of cache capability is consistent with the theoretical calculation based on the port cache size.
