A Throughput-Centric View of the Performance of Datacenter Topologies

This repository contains the code for computing throughput upper bound (TUB). TUB is tight and scalable metric for estimating worst-case throughput of uni-regular and bi-regular topologies, and very useful to testing whether a designed topology is non-blocking or not (ie, can support every traffic matrix without the need to drop any traffic).

Instructions

Code Structure

metric/ code for computing throughput upper bound.
topo_repo/ contains a set of topologies (Jellyfish, Xpander, FatClique, Clos)
utils/ code for loading/storing the topologies and generating maximal permutation traffic matrix.

Installation

(1) Clone this repo

$ git clone https://github.com/USC-NSL/TUB.git

(2) Python version. This repo is tested with the following two combinations;

Python 2.7.17 + pip 20.2.3 (used to generate results of the paper)
Python 3.7.10 + pip 21.1.1

(3) Install dependencies

$ pip install -r requirements.txt

Required dependencies:

networkx
scipy
numpy
python-igraph
matplotlib

How to use?

There are two ways of using this repository;
(1) Storing the topology and computing the TUB using main_tub.py.

$ python main_tub.py path_to_topology_file

In order to use this approach, you should store your topology in the format of Topology class provided under topo_repo/topology.py. After you initialized the class, you can call store_model(model, file_path) from utils/utilities.py to store the topology.

Examples. For each topology family used in the paper (Jellyfish, Xpander, FatClique, Clos), we have provided a few examples in topo_repo/ using which you can try TUB. For example, to compute TUB on Jellyfish with 32K servers and 8 servers per switch, you can try the following.

$ python main_tub.py topo_repo/Jellyfish/jf_R_32_H_8_N_32000

(2) Direct function call.
Alternatively, you can call get_throughput_upper_bound() from metric/tub.py in your code. The function gets three inputs and computes the TUB. Please visit metric/tub.py for more information.

Note. Depending on your computing resources, you might face out-of-memory error when trying to compute the TUB on some of the provided examples.

Using a 64GB machine, we are able to compute TUB for all the examples
Using a 32GB machine, we can compute TUB for all except Jellyfish and Xpander with ~128K servers.