[Feature Request]: Heterogeneity and Switch Support

[liangyuRain]

Feature Request

Hi there! I came across your paper on arXiv and noticed that TACOS supports schedule generation for heterogeneous and switch topologies. However, I couldn’t find the features implemented in this repo. Could you kindly provide guidance or support for this? Thank you!

Project Version

No response

Miscellaneous

No response

[xuruifan]

I found that this paper was published in the MICRO conference and the artifact has already been done. I'm really interested in the topology-related code tested in the paper. Could you please make it publicly available?

I also have a question regarding how to use the simulator backend with congestion-aware supports. Could you provide some guidance or documentation on this? Thank you!

[beatlesforever]

Hi! I also wish to access the original code and am currently facing issues during compilation. I was wondering if you encountered any problems as well? Any guidance would be greatly appreciated. Thank you!

[sabujlaskar]

@liangyuRain I was also looking for switch support. Although it is not available in the repository, we can represent the switch with point-to-point connections as shown in the paper with switch unwinding, and implement the topology accordingly.

[beatlesforever]

@liangyuRain I was also looking for switch support. Although it is not available in the repository, we can represent the switch with point-to-point connections as shown in the paper with switch unwinding, and implement the topology accordingly.

@sabuj7177 I would be immensely grateful if you could kindly provide an example code. Thank you so much in advance! 🙏

[sabujlaskar]

Hi @beatlesforever,

I am sharing my understanding of switch unwinding as explained in the TACOS paper. Consider a 3_3 SW_SW topology with a bandwidth of 1200_900 GBps. In this setup, nodes 0, 1, and 2 are connected to one switch, and nodes 3, 4, and 5 are connected to another, and so forth. By unwinding this topology with a degree of 2, it becomes a fully connected topology where node 0 connects to both nodes 1 and 2, with each channel having a bandwidth of 600 GBps.

This understanding is based on the TACOS paper. For other switches, a similar approach can be applied. Writing the example code should be straightforward; you need to create a topology file and establish connections between nodes. For the switch connecting nodes 0, 1, and 2, the code would look like this:

connect(0, 1, 600, latency, bidirectional=True) # 600 because the total BW in the first dimension is 1200
connect(0, 2, 600, latency, bidirectional=True)
connect(1, 2, 600, latency, bidirectional=True)

You will need to write similar code for all the switches. This suggestion is based on my understanding of the TACOS paper and their codebase. Please confirm with the author before using this setup.

[linsj20]

We are now working on a general switch representation that avoids switch unwinding. Thank you for your attention!