Cloud Infrastructure: Data Center Network Topologies | Assignments Network Design

14-848: Cloud Infrastructure

Project #1: Data Center Network Topologies

Overview

This lab asks you to simulate three different network topologies, measuring the effective throughput of each

under various workloads, and determining the pattern of under- and over-subscribed regions for each work

load. The results of the analysis should be presented meaningful and accessibly in each of numerical summaries,

plotted summaries, and a color coded “heat map” diagrams of the network.

Group Size

1-2 people. Divide-and-conquer is not permitted. Teams must work collaborative on the entire project.

Tools

• Python or Java

• MiniNet with WireShark, etc

• GNUPlot, Excel or another plotting utility

• Excel, code, or some other tool for studying logs or other simulated network data/observations

• Draw.io, Creately, or any other drawing/diagramming utility

Network Topologies

You should simulate each of the following topologies:

• 3-tier with redundant core (2xCore + Aggregate + Access + Hosts, each core connected to each

aggregate)

• Fat-Tree (As discussed in class)

• 2-layer spine-leaf (Clos-like with each leaf-level switch connected to each core router)

Network size:

• This is up to you, but the networks need not be large to be sufficiently large. Consider the number of

ports/switch, the fan-out, and the workloads. Guesstimate how large the network needs to be to have

the features you need to demonstrate its behavior. Test this configuration. Then, try something with

more fan-out, maybe twice as much. Are the big-picture results stable? If so, you are probably good. If

not, try to understand why not, take another guess, and try again. If you want to be particularly

thorough and convincing, you can run your simulation in an automated way with different network

sizes, plotting key results, and demonstrating the stability.

Workloads

Each simulated network should be tested with workloads as follows:

• Single-pair, East-west (peer-to-peer)

• Multiple pair, east-west (peer-to-peer)

• Multiple pair, Pod-local and/or top-or-rack local

• Incast (concurrent results percolating back up)

• One or more workload you consider interesting for each topology

Workloads can be achieved by real, meaningful client-server traffic, e.g. you can set up clients and servers, or by

setting up workload generators that generate traffic in a distribution that meaningfully approximates such real

clients.

Partial preview of the text

Download Cloud Infrastructure: Data Center Network Topologies and more Assignments Network Design in PDF only on Docsity!

14 - 848: Cloud Infrastructure Project #1: Data Center Network Topologies Overview This lab asks you to simulate three different network topologies, measuring the effective throughput of each under various workloads, and determining the pattern of under- and over-subscribed regions for each work load. The results of the analysis should be presented meaningful and accessibly in each of numerical summaries, plotted summaries, and a color coded “heat map” diagrams of the network. Group Size 1 - 2 people. Divide-and-conquer is not permitted. Teams must work collaborative on the entire project. Tools

Python or Java
MiniNet with WireShark, etc
GNUPlot, Excel or another plotting utility
Excel, code, or some other tool for studying logs or other simulated network data/observations
Draw.io, Creately, or any other drawing/diagramming utility Network Topologies You should simulate each of the following topologies:
3 - tier with redundant core (2xCore + Aggregate + Access + Hosts, each core connected to each aggregate)
Fat-Tree (As discussed in class)
2 - layer spine-leaf (Clos-like with each leaf-level switch connected to each core router) Network size:
This is up to you, but the networks need not be large to be sufficiently large. Consider the number of ports/switch, the fan-out, and the workloads. Guesstimate how large the network needs to be to have the features you need to demonstrate its behavior. Test this configuration. Then, try something with more fan-out, maybe twice as much. Are the big-picture results stable? If so, you are probably good. If not, try to understand why not, take another guess, and try again. If you want to be particularly thorough and convincing, you can run your simulation in an automated way with different network sizes, plotting key results, and demonstrating the stability. Workloads Each simulated network should be tested with workloads as follows:
Single-pair, East-west (peer-to-peer)
Multiple pair, east-west (peer-to-peer)
Multiple pair, Pod-local and/or top-or-rack local
Incast (concurrent results percolating back up)
One or more workload you consider interesting for each topology Workloads can be achieved by real, meaningful client-server traffic, e.g. you can set up clients and servers, or by setting up workload generators that generate traffic in a distribution that meaningfully approximates such real clients.

Deliverables

Source code implementing the network simulation (topologies) and host simulation (traffic generation)
A report containing the following information: o Description of the organization of your source code and other deliverables o Description of any assumptions you made and why o Description of any requirements that were not completed or only partially completed o A description of the workload(s) you defined, and an explanation of why you considered them interesting. o Prose describing what you observed w.r.t. the behavior of the various topologies o A network diagram showing each simulated topology and host o A numerical summary of the observed results. This should be where, for each case, you provide descriptive statistics, e.g. means, medians, etc. There is no specific requirement for what measurements or statistics should be provided. It is up to you to determine how to best convey what you learned. o Plots, charts, and/or other visualizations, e.g. histograms, line plots, bar charts, etc, to help the reader understand and build intuition about the numerical summaries you provide, as well as anything else you think it might be helpful to express this way. There is no specific requirement for what visualizations should be provided. It is up to you to determine how to best convey what you learned. o For each workload, a network diagram showing each simulated topology and host color coded as a “heat map” to show utilization. Questions? Concerns? Ideas? We’re here to help! Please ask!

Cloud Infrastructure: Data Center Network Topologies, Assignments of Network Design

Related documents

Partial preview of the text

Download Cloud Infrastructure: Data Center Network Topologies and more Assignments Network Design in PDF only on Docsity!