Task perception based transmission control method for data center network

Abstract

The invention discloses a task perception based transmission control method for a data center network. The method includes steps: a sending end in a communication process calculates current network available bandwidth according to explicit feedback information and adjusts a sending rate according to the current network available bandwidth and a sent flow trailing degree, a router feeds back network loads through packet header ECN information, and a receiving end redisplays the ECN information and calculates predicted task completion time; the sending end calculates the trailing degree according to the predicted task completion time carried in ACK by the receiving end and predicated flow completion time calculated by the sending end, and the tailing degree and the explicit feedback information jointly serve as bases for sending rate adjustment. By combination of end-to-end task information and network congestion information feedback, extra control cost is avoided, deployment difficulty is lowered, protocol universality is improved while short task completion time and high throughput rate are guaranteed, and accordingly waiting time after request sending is shortened for users.

Description

technical field [0001] The invention relates to a task-aware transmission control method in a data center network. Background technique [0002] In modern data centers, tree-based divide-and-conquer algorithms are commonly used to deliver data or calculate to thousands of servers. After each server calculates or processes, the results will be transmitted to the aggregation machine (Aggregator ) for combination and processing, and the final result can only be returned to the user after the processing by the aggregator is completed. Such parallel processing can greatly improve operating efficiency, reduce the waiting time of user requests, and improve and enhance user experience. However, it is obvious that if some servers process the delivered work very slowly, the aggregator must wait until all the data arrives before starting to work, resulting in longer waiting time for users. This is known as the straggler phenomenon. Recent studies have shown that network latency acco...

AI Technical Summary

Problems solved by technology

And if these small streams need to spend an additional RTT or even longer time to wait for the scheduling decision of the central scheduler, their completion time will increase by at least 100%.

[0005] It can be seen that in the data center network, the existing transmission control methods, or use a huge system to ensure precise control, but increase the calculation overhead and transmission delay, and at the same time, due to large changes to the existing data center network, lose or use flexible deployment and simple mechanisms to ensure fast response, but the results are often inaccurate, and there is no way to reduce task completion time and user waiting time

Images