Tip:
Highlight text to annotate it
X
NoC Hardware System Performance Analysis Tool Demo - 2
This video demonstrates how to use the EMBRACE NoC monitor. Six traffic patterns are chosen as examples.
Connect to remote server and download the bit file to FPGA
Open the remote server and download the bit file to FPGA
USB-Blaster is the usb-blaster connected to host computer. Programming FPGA takes a while.
Use the monitoring software to analyze the NoC performance
Open the software, configure the NoC system size and run time, then open the Uart port.
The software receives the data from HW NoC system.
Click 'Traffic map' to generate the traffic heat map
Now it is traffic pattern (a), i.e. (8,*) sends packets to (*,8). It can be seen that the software plots the traffic heat map successfully.
We can also watch the average throughput, the number of lost/corrupted packets.
The throughput of each nodes can be displayed in a 3D throughput diagram.
Now the NoC system is changed to traffic pattern (b) with varied Packet Injection Rate (PIR). The traffic patterns can be selected using switches on the DE4 board.
The 3D throughput diagram can be rotated to any angles; then the throughput can be compared visually.
Now change to traffic pattern (c). A traffic congestion occurs at the west output channel of node (3,2).
EDAR is an adaptive routing algorithm with congestion-avoidance. The packets are forwarded to both west and south directions.
The throughput of node (3,2) and (2,3) is higher than other nodes.
Now change to traffic pattern (d). West channel of (6,2) is faulty.
EDAR bypasses the faulty channel and uses another fault-free channel to transmit packets.
Now change to traffic pattern (e). WC(6,2), WC(5,3~6) are faulty.
EDAR bypasses the faulty channel and uses fault-free channels to transmit packets.
Traffic pattern (f) with clustered faulty channels.
Faulty channels in the HW NoC are displayed in black colour.
EDAR does not have look-ahead function; therefore the packets are forwarded to node (3,2) and lost or corrupted.
Current fault model is Stuck-at model and the number of corrupted packets is 255.
If the fault model is crosstalk model; then the number of lost/corrupted packets is ~25/230.
Therefore, this software can be used to analyse the hardware NoC system performance, especially the fault-tolerance capability.