NS-3 (Discrete-Event Network Simulator)

NS-3 Tutorial Notes

ASTRA-sim with NS-3 as Network Backend

1 - Generating Chakra ETs as Input

2 - Running ASTRA-sim+ns-3

5 - Other Tests

Problems Found!

Check out the following to know how ASTRA-sim and ns3 work together (what signals are sent)

3 - Interface of ASTRA-sim+ns-3

Check out the following to get an insight of how ASTRA-sim+NS3 scheduling works:

4 - Scheduling in ASTRA-sim + NS-3

TLDR: ASTRA-sim and ns-3 sequentially process the ETs, and does not show any network fidelity. For example, there is no implementation of barrier for collective communications, which simply processes collective nodes as soon as they are scheduled (without considering synchronization with ranks).

• Child Problem: Is this really a fundamental problem of ASTRA-sim or just an implementation problem?
• Child Problem: What defines network fidelity?

TODO:

1. How is duration micro used for collective operations
   1. collective operations do not use duration_micros
   2. problem is that collective operations are simply defined by link latency and transmission latency
   3. ASTRA-sim does not consider the pre-posting of RECV before SEND is sent or the control messages between SEND/RECV
      1. ASTRA-sim only takes care of this in the system perspective but does not provide coordination between nodes before a collective operation is called
2. How can ASTRA-sim or NS-3 be modified to consider barrier or synchronization among nodes?
   1. A global variable could be made within ns-3 to keep track of send and recv packets
      1. e.g. send only happens once there is a corresponding recv