Receiver-Driven Dynamic Pull Mode for Datacenter Congestion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In network transport protocols, managing incast traffic at a receiver with high latency rendezvous-based network handshakes can be expensive, especially in networks with large radii and varying round trip times, requiring apriori knowledge of message size and network conditions, which is not feasible in under-provisioned networks.
Innovation Solution
Implementing a dynamic pull mode control system where a network endpoint receiver monitors packet traffic characteristics and switches from push mode to pull mode, sending pull mode request packets to control transmitter pacing, allowing for adaptive packet transmission rates without requiring knowledge of network round trip time or message size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If high latency rendezvous based network handshakes are used to manage incast traffic, then packet flow control is achieved, but system complexity and latency increase significantly
Solution Approach 1:
The patent extracts the flow control decision-making from the complex rendezvous handshake process and places it directly at the receiver end. The receiver independently monitors its buffer status and sends simple pause/resume signals to transmitters, eliminating the need for complex coordinated handshakes while achieving reliable packet flow control.
Solution Approach 2:
Instead of the transmitter controlling packet flow unilaterally through complex handshakes, the patent inverts the approach by enabling the receiver to control flow based on its own buffer conditions. The receiver sends pause frames when buffers are full and resume frames when ready, reversing the traditional control paradigm and simplifying the overall system.
2Reliability
If apriori knowledge of message size and RTT is required for rendezvous handshakes, then flow control accuracy is improved, but adaptability to network variations deteriorates
Solution Approach 1:
The patent implements a feedback mechanism where the receiver continuously monitors its buffer occupancy and sends pause/resume signals to transmitters based on actual conditions. This closed-loop feedback eliminates the need for apriori knowledge of message size and RTT, allowing the system to adapt dynamically to network variations while maintaining accurate flow control.
Solution Approach 2:
The system transitions from static apriori parameters to dynamic adaptive control. The receiver's pause and resume decisions are made in real-time based on current buffer status, enabling the system to adapt to varying network conditions, message sizes, and RTT without requiring预先 knowledge of these parameters.
3Productivity
If push mode transmission is used initially, then transmission speed is maximized, but receiver buffer overflow and packet loss increase
Solution Approach 1:
The receiver monitors its buffer occupancy in real-time and sends pause frames to transmitters when buffers are full, preventing packet overflow. This feedback mechanism allows the system to maintain high transmission speed during normal operation while preventing packet loss when buffers are saturated, balancing productivity with harm prevention.
Solution Approach 2:
The system operates in periodic cycles of high-speed push mode transmission interspersed with pause periods when buffers are full. The receiver sends resume frames to re-establish transmission when buffers are cleared, creating a rhythmic pattern of active transmission and pause that maintains both speed and prevents overflow.
Data Source
AI summary
A network endpoint receiver controls packet flow from a transmitter. Packets are received via a network in packet traffic according to a push mode, where the transmitter controls pacing of transmitting the packets. Characteristics related to the packet traffic are monitored at the receiver. The monitored characteristics are compared to reception performance parameters, and based on the comparison, a decision is made to switch from the push mode to a pull mode for controlling the packet flow. The receiver transmits a pull mode request packet to the transmitter, where the pull mode request packet indicates a pacing of subsequent packets transmitted by the transmitter to the receiver in accordance with the pull mode. Pacing of further transmitted packets may be controlled by subsequent pull mode request packets sent over time to the transmitter by the receiver. Similarly, the receiver may control additional transmitters to transmit at equal or different rates.


