Exploring XON-XOFF Flow Control with Modeling and Simulation
Below is a browsable image of the model.
- For an executable version,
- To view or save the MoML file for this model, click here.
Introduction
Purpose of this demonstration system is to explore XON-XOFF Flow Control using Modeling and Simulation. XON-XOFF Flow Control is used to control the data transfers between computers and other devices in an asynchronous serial connection. In this simulation model we have two Ingress queues representing two master devices and a single Egress Queue representing a Slave device. Block diagram of the system is shown below
A set of transactions from different channels (Ch1, Ch2,...Chn) arrives at Input Ingress Buffer and the Ingress Buffer arbiter will send packets to individual output Q or Egress Buffer. In the current model, we assume the mapping of input to the final Output Q is pre-provision. (From the above Input Ch1 map to OutputCh1. ) The Flow Control between Output Ch1 and Input Ch1 can be Xon/Xoff based or Credit Based. Multiple Input Q will flow through a inter-mediate Channel. The flow control of these channel will be Xon/ Xoff Based.
(Note : The max. traffic rate, and the drain rate of any channel should be parameterised, and the actual drain rate will be Flow control back-pressured based.)
Flow Control Logic is
defined using VisualSim Scripting language, transactions will be sent
out of Ingress Buffer only if Egress Buffer has sufficient
space.
The model is broken down into the following sections
1. Traffic source: The source has a input rate with a uniform distribution.
2. Setting up the forward and output queues
3. Writing to the forward queue
4. Removing from the output queues
5. Round-robin algorithm
6. Applying Backward Pressure
7. Reading from the forward queue and placing them in the output queue along with the estimated processing time in the output queue: The process of transferring the data from the forward queue to the output queue is done once the backward pressure has been evaluated and the forward queue is free to send the data. This is done by first removing the data from the Forward Queue, computing processing delay at the output queue and then placing the frames in the associated output port queue along with the processing time.
8. Statistics has been gathered on the forward and output queues, and the latency of each frame at the exit of the output queue. The latency is computed by calculating the time from the packet generation to the exit of the output queue (“DELTA = TNow – TIME”).
Analysis
Two results are captured in this model. The first is the statistics of the forward and output queue occupancy while the second is the end-to-end latency of the each frame. There are considerable spikes in the latency curve indicating the unpredictable nature of this algorithm.This indicates that the design did not scale well. The forward queues are somewhat overdesigned. Also, the back pressure is never triggered as the data is not arriving fast enough. The round-robin algorithm can be improved using a weighted average by adding more knowledge about the content of the queues.