thanks
the time it takes (clocks) for the packet to arrive varies according to network congestion
All you can measure is the latency and throughput of your design. The throughput would be theoretical maximum.
Ok, but how i can measure them ????
You know what your clock speed is.
You know what the bus width is.
So therefore you can calculate the max bandwidth in bit/s
If you cannot process each word on every clock - then that is a limitation. You should either know this limitation, or you can measure it on hardware.
If this is ethernet - then overload the system with more data than it can handle. Ensure all packets are the same size, then counter the number of packets over a given time window (say 1second) and then you'll have your bandwidth
Latency throughput and bandwidth can only be specified by you, as the designer, as you know the specification of the system (the design tools do not). Energy use has to be measured with an ammeter.
This is in the case of a normal network designed in hardware, not on the chip based VHDL
After all thank you very much
The analysis is the same in either case, not sure why you think it would be any different (it's a network in both cases). As the packet arrival times are random in nature (based on your previous posts) then the entire analysis is statistical and will end up with a curve of possible latencies through your network on chip. Changes in the loading of the network will affect the latency as will network congestion.
The easiest analysis for worst case latency is to just assume all network resources want the same destination and run a simulation of the system with all ports requesting the same sink port all at the same time. The port that takes the longest to be serviced by the arbiter is the worst case latency in an un-congested network. If there is any buffering in the network then having all the ports filled with data before doing this and seeing when the last packet from the last serviced port is complete will give you the worst case latency in a congested system. Some of these calculations can be very simple if the arbitration is simple and the number of clocks between servicing ports is a consistent value.
If you need a more comprehensive analysis you'll have to resort to determining the service time of the arbiter then use that data as input to something like matlab to analyze the port requests using a statistical approach. I'm not any kind of expert in queuing theory, so I can't help you with setting up such an analysis, but there are a lot of papers written up on that kind of analysis.
We use cookies and similar technologies for the following purposes:
Do you accept cookies and these technologies?
We use cookies and similar technologies for the following purposes:
Do you accept cookies and these technologies?