14-760: Adv. Real-World Networks Switching fabric,
Description: 14-760: Adv. Real-World Networks Switching fabric, Routers, Routing Lecture 5 Spring 2020 (Kesden) Remember Socrative https:api.socrative.comrcFX5GFK Where are we? Last class Reviewed the basics of networking. Hopefully a refresher
Related Topics
Download Presentation
"14-760: Adv. Real-World Networks Switching fabric," is the property of its rightful owner. Permission is granted to download and print the materials on this website for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Presentation Transcript
slide1. 14-760: Adv. Real-World Networks Switching fabric, Routers, Routing
Lecture 5 * Spring 2020 (Kesden)<br>
slide2. Remember Socrative https://api.socrative.com/rc/FX5GFK<br>
slide3. Where are we? Last class
Reviewed the basics of networking. Hopefully a refresher from prior class(es)
Today
Examine the role of switches and routers vis-à-vis data center and Internet design
Much of this will likely be a review for some (but not nearly all)
Next class
Look at the internals architecture of switches and routers.
This is likely totally new to most.<br>
slide4. Today Definitions
Data Center Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide5. Definitions for Today “Routing” (Network-layer packet switching)
The hop-to-hop management of the movement of traffic among networks
“Switching” (Link-layer frame switching)
The movement of frames within a local network
Note:
There is nothing wrong with using the word “switching” at other-than the link layer, e.g. network layer switch, application layer switch, etc.
Such references describe switches that make decisions based upon information present at other layers of the network stack
That just doesn’t happen to be the way I am using the words today, so I define them here for today’s discussion.<br>
slide6. Today Definitions
Data Center Challenges and Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide7. DC Networking Challenges Within a Data Center:
Distances are short, the speed of light is less of a factor
Serialization/Deserialization delay (bits/second) is a factor
Managing bandwidth becomes complex as it is easy to add wire, but hard to use it.
vs. Long haul:
Speed of light is significant source of latency, unlikely to be resolved soon
Additional bandwidth can typically readily be achieved with parallelism, e.g. more fiber<br>
slide8. Types of Traffic Ingress – Traffic entering from the outside, e.g. requests entering a data center
Most commonly from outside requests.
Can be from queries to related off-site services.
Egress – Traffic leaving to the outside, e.g. responses, streaming video, database results, etc.
Most commonly from providing service.
Can be queries to related off-site services
East-West – Traffic within a data center that crosses higher level switches
Most commonly from communication within integrated services
Incast – Traffic within a data center that becomes concentrated
Most commonly from parallel query concentrating results<br>
slide9. Types Of Traffic: Ingress Core Aggregation Leaf<br>
slide10. Types Of Traffic: Egress Core Aggregation Leaf<br>
slide11. Types Of Traffic: East-West Core Aggregation Leaf<br>
slide12. Types Of Traffic: IncasT Core Aggregation Leaf<br>
slide13. Today Definitions
Data Center Challenges and Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide14. Network Switches Ports connect to network segments
Hosts and uplink at leaf level
Other switches at other levels
Switches have to move messages (frames, packets, etc) from one port to another
Hardware solutions: Crossbar, etc
Shared memory solutions: input and output queues
Limited throughput (processor time, shared memory, crossbar, etc)
Limited number of ports, limited throughput per port, limited aggregate throughput
Commodity-level capacities are relatively affordable
Trying to buy more capacity is expensive with limited gain<br>
slide15. Since, beyond a certain point, we can’t make switches wider and/or faster, we need to “fan out”, most commonly with a tree topology
Venerable 3-tier network is a straight-forward example: DC Topology: Venerable 3-Tier Core Aggregation Leaf<br>
slide16. Can add a redundant core for increased throughput and resilience DC Topology: Venerable 3-Tier With Redundant Core Core Aggregation Leaf<br>
slide17. Scales nicely, but …
Higher up gets over-subscribed since everything passes through
Over-subscription increases with scale
Request-to-stream and host-to-host cases generate bottlenecks DC Topology: Venerable 3-Tier, Analysis 1-2 Switch Throughput W Switch Throughput W 2 Switch Throughput<br>
slide18. DC Topology: Clos Network By Piggly (talk) (Uploads) - Transferred from en.wikipedia to Commons.,
Public Domain, https://commons.wikimedia.org/w/index.php?curid=61536102 Allocating an input port, and associated throughput,
Allocates path whole way through.
NxN connectivity with switches with less than NxN connectivity
Basically a way to make a large NxN switch
An expensive solution and not likely to need all throughput capacity simultaneously<br>
slide19. DC Topology: Leaf and Spine Type of Clos network
Essentially folded, but still N-to-N connections
Derived from old phone company architecture,
invented in 1950s.
Sometimes Called “Folded Clos”
All paths are same length from edge to edge
Need to pick path, as can choose any middle router
Very redundant
Can implement at layer-2 or layer-3<br>
slide20. DC Topology: Fat-Trees With Skinny Switches: Goals Use all commodity switches
Full throughput from host-to-host
Compatible with usual TCP/IP stack
Better energy efficiency per unit throughput from more smaller switches than fewer bigger switches
As you look at the next slide note that, given K ports/switch, we are using:
K/2 ports up + K/2 ports down, or
K ports up, or
K ports down<br>
slide21. Fat Tree (K=4) K Ports/switch Note the replacement of aggregation layer switches with 2 layers of K/2 K-port switches (K/2)2 core routers K pods K-port switches support K3/4 servers (K/2)2 servers per pod Base Image Credit: Mysore, et al, "PortLand: a scalable fault-tolerant layer 2 data center network fabric", SIGCOM, October 2009, Pages 39-50.<br>
slide22. DC Topology: Fat Tree Details K-ary fat free: three layers (core, aggregation, edge)
Each pod consists of (K/2)2 servers and 2 layers of K/2 K-port switches.
Each edge switch connects (K/2) servers to (K/2) aggregator switches
Each aggregator switch connects (K/2) edge and (K/2) core switches
(K/2)2 core switches, each ultimately connecting to K pods
Providing K different roots, not 1. Trick is to pick different ones
K-port switches support K3/4 servers/host:
(K/2 hosts/switch * K/2 switches per pod * K pods)<br>
slide23. DC Topology: What is Actually Used? Redundant 3-Tier architectures
Often used for general IT data centers
They scale well
Clos Networks
Too expensive
Not likely to be sufficiently utilized
Fat Tree Networks
Look good on paper and very practical
Might use them, if we needed to
Leaf-And-Spine
Right now, port density seems to be supporting their use
Not clear that next generation of switches won’t trade bandwidth for port density and push us back to Fat Tree or other<br>
slide24. DC Topology: What Is Actually Used? Note that there are many other possibilities, e.g.
Grid/Cube-like architectures: Mesh, Torus, Hypercube, Butterfly, Dragonfly, HyperX, etc
Note also that different parts of a data center, with different requirements may be networked differently, e.g. commodity file storage, vs bulk streaming, vs app servers, etc.<br>
slide25. Today Definitions
Data Center Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide26. Parallelism Within Networks: Traditional Thinking Parallelism is for redundancy
Run self-healing (periodically updating)a Spanning Tree Protocol (STP) among switches to agree a network tree without cycles
Avoid frames being forwarded forever in loops
Uses redundancy to select an alternate path in the event of hard failure<br>
slide27. Parallelism Within Networks: Bandwidth-Motivated Thinking Parallelism is for bandwidth
Have parallelism where needed to support traffic patterns
Find ways to use bandwidth without causing problems
Avoid looping
Use all available paths<br>
slide28. Parallelism Within Networks: Where Is Bandwidth Needed? Depends upon usage pattern
If most activity is rack-local
At the rack switch
3-Tier might be fine
If most of the activity is east-west
Need more capacity at higher level(s), e.g. core and/or aggregation, etc
If most of activity is concentrating at egress (or ingress)
Need more core upstream capacity (supported by core)<br>
slide29. Parallelism Within Networks: How To Use Parallelism? Goals:
Must avoid cycles
Usage should be divided among paths
Ways to divide usage
Each use, e.g. network session, picks different path
Uses divide their traffic among parallel paths<br>
slide30. Using Parallelism Within Networks: Solvable at Different Levels (Examples) MPTCP – TCP that shares multiple paths for single session
Creates many sub-sessions and multiplexes overarching session among them
PORTland network – Link-Layer solutions
Path is based upon sender’s location in network<br>
slide31. Today Definitions
Data Center Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide32. Routing As we leave the domain of a single network, we enter the domain of internetworking
Many techniques for global connectivity are possible, but “packet switching” is the most common for Internet traffic as it crosses providers
Other types of switching, including (and perhaps especially) virtual circuits, e.g. ATM, are commonly used within provider networks.
Circuit switching is a fading technology, e.g. POTS, ISDN.
Packet switching involves a “router” making hop-by-hop decisions about how to “route” a packet.<br>
slide33. Routers May Also be called “Internet Switches” or “Network-layer switches”
May also perform other roles
Rule-based Firewalls
Rule-based load-balancing
Network Address Translation (NAT)
Virtual Private Networks and Tunneling<br>
slide34. Routing Protocols Routers exchange information with each other about the state of the network
Routers collect this into routing tables
Algorithms reduce the routing tables to simple forwarding tables that the switch can quickly consult w.r.t. which port should be the destination port for a particular incoming packet.<br>
slide35. Data Plane Vs Control Plane Routing algorithms run in the Control Plane
The control plane decides the rules for forwarding, but it does not do the forwarding
The Data Plan is where the forwarding actually occurs
Packets move in and out via the data plane, which consults the distilled tables provided by the control plane<br>
slide36. Today Definitions
Data Center Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide37. Next class Routers and switches need to sustain a very, very high throughput.
They need to very rapidly
Source a packet
Decide what to do with it
Sink it to another port
This could involve buffering it at the input port, for a decision, and/or at the output port
This could involve a computer-like shared memory system or a more purpose-built hardware switch<br>
slide38. Next class, cont. Next class, we are going to take a look a the architectures, where the points of contention are within them, discuss their performance and relative costs, &c.<br>
Lecture 5 * Spring 2020 (Kesden)<br>
slide2. Remember Socrative https://api.socrative.com/rc/FX5GFK<br>
slide3. Where are we? Last class
Reviewed the basics of networking. Hopefully a refresher from prior class(es)
Today
Examine the role of switches and routers vis-à-vis data center and Internet design
Much of this will likely be a review for some (but not nearly all)
Next class
Look at the internals architecture of switches and routers.
This is likely totally new to most.<br>
slide4. Today Definitions
Data Center Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide5. Definitions for Today “Routing” (Network-layer packet switching)
The hop-to-hop management of the movement of traffic among networks
“Switching” (Link-layer frame switching)
The movement of frames within a local network
Note:
There is nothing wrong with using the word “switching” at other-than the link layer, e.g. network layer switch, application layer switch, etc.
Such references describe switches that make decisions based upon information present at other layers of the network stack
That just doesn’t happen to be the way I am using the words today, so I define them here for today’s discussion.<br>
slide6. Today Definitions
Data Center Challenges and Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide7. DC Networking Challenges Within a Data Center:
Distances are short, the speed of light is less of a factor
Serialization/Deserialization delay (bits/second) is a factor
Managing bandwidth becomes complex as it is easy to add wire, but hard to use it.
vs. Long haul:
Speed of light is significant source of latency, unlikely to be resolved soon
Additional bandwidth can typically readily be achieved with parallelism, e.g. more fiber<br>
slide8. Types of Traffic Ingress – Traffic entering from the outside, e.g. requests entering a data center
Most commonly from outside requests.
Can be from queries to related off-site services.
Egress – Traffic leaving to the outside, e.g. responses, streaming video, database results, etc.
Most commonly from providing service.
Can be queries to related off-site services
East-West – Traffic within a data center that crosses higher level switches
Most commonly from communication within integrated services
Incast – Traffic within a data center that becomes concentrated
Most commonly from parallel query concentrating results<br>
slide9. Types Of Traffic: Ingress Core Aggregation Leaf<br>
slide10. Types Of Traffic: Egress Core Aggregation Leaf<br>
slide11. Types Of Traffic: East-West Core Aggregation Leaf<br>
slide12. Types Of Traffic: IncasT Core Aggregation Leaf<br>
slide13. Today Definitions
Data Center Challenges and Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide14. Network Switches Ports connect to network segments
Hosts and uplink at leaf level
Other switches at other levels
Switches have to move messages (frames, packets, etc) from one port to another
Hardware solutions: Crossbar, etc
Shared memory solutions: input and output queues
Limited throughput (processor time, shared memory, crossbar, etc)
Limited number of ports, limited throughput per port, limited aggregate throughput
Commodity-level capacities are relatively affordable
Trying to buy more capacity is expensive with limited gain<br>
slide15. Since, beyond a certain point, we can’t make switches wider and/or faster, we need to “fan out”, most commonly with a tree topology
Venerable 3-tier network is a straight-forward example: DC Topology: Venerable 3-Tier Core Aggregation Leaf<br>
slide16. Can add a redundant core for increased throughput and resilience DC Topology: Venerable 3-Tier With Redundant Core Core Aggregation Leaf<br>
slide17. Scales nicely, but …
Higher up gets over-subscribed since everything passes through
Over-subscription increases with scale
Request-to-stream and host-to-host cases generate bottlenecks DC Topology: Venerable 3-Tier, Analysis 1-2 Switch Throughput W Switch Throughput W 2 Switch Throughput<br>
slide18. DC Topology: Clos Network By Piggly (talk) (Uploads) - Transferred from en.wikipedia to Commons.,
Public Domain, https://commons.wikimedia.org/w/index.php?curid=61536102 Allocating an input port, and associated throughput,
Allocates path whole way through.
NxN connectivity with switches with less than NxN connectivity
Basically a way to make a large NxN switch
An expensive solution and not likely to need all throughput capacity simultaneously<br>
slide19. DC Topology: Leaf and Spine Type of Clos network
Essentially folded, but still N-to-N connections
Derived from old phone company architecture,
invented in 1950s.
Sometimes Called “Folded Clos”
All paths are same length from edge to edge
Need to pick path, as can choose any middle router
Very redundant
Can implement at layer-2 or layer-3<br>
slide20. DC Topology: Fat-Trees With Skinny Switches: Goals Use all commodity switches
Full throughput from host-to-host
Compatible with usual TCP/IP stack
Better energy efficiency per unit throughput from more smaller switches than fewer bigger switches
As you look at the next slide note that, given K ports/switch, we are using:
K/2 ports up + K/2 ports down, or
K ports up, or
K ports down<br>
slide21. Fat Tree (K=4) K Ports/switch Note the replacement of aggregation layer switches with 2 layers of K/2 K-port switches (K/2)2 core routers K pods K-port switches support K3/4 servers (K/2)2 servers per pod Base Image Credit: Mysore, et al, "PortLand: a scalable fault-tolerant layer 2 data center network fabric", SIGCOM, October 2009, Pages 39-50.<br>
slide22. DC Topology: Fat Tree Details K-ary fat free: three layers (core, aggregation, edge)
Each pod consists of (K/2)2 servers and 2 layers of K/2 K-port switches.
Each edge switch connects (K/2) servers to (K/2) aggregator switches
Each aggregator switch connects (K/2) edge and (K/2) core switches
(K/2)2 core switches, each ultimately connecting to K pods
Providing K different roots, not 1. Trick is to pick different ones
K-port switches support K3/4 servers/host:
(K/2 hosts/switch * K/2 switches per pod * K pods)<br>
slide23. DC Topology: What is Actually Used? Redundant 3-Tier architectures
Often used for general IT data centers
They scale well
Clos Networks
Too expensive
Not likely to be sufficiently utilized
Fat Tree Networks
Look good on paper and very practical
Might use them, if we needed to
Leaf-And-Spine
Right now, port density seems to be supporting their use
Not clear that next generation of switches won’t trade bandwidth for port density and push us back to Fat Tree or other<br>
slide24. DC Topology: What Is Actually Used? Note that there are many other possibilities, e.g.
Grid/Cube-like architectures: Mesh, Torus, Hypercube, Butterfly, Dragonfly, HyperX, etc
Note also that different parts of a data center, with different requirements may be networked differently, e.g. commodity file storage, vs bulk streaming, vs app servers, etc.<br>
slide25. Today Definitions
Data Center Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide26. Parallelism Within Networks: Traditional Thinking Parallelism is for redundancy
Run self-healing (periodically updating)a Spanning Tree Protocol (STP) among switches to agree a network tree without cycles
Avoid frames being forwarded forever in loops
Uses redundancy to select an alternate path in the event of hard failure<br>
slide27. Parallelism Within Networks: Bandwidth-Motivated Thinking Parallelism is for bandwidth
Have parallelism where needed to support traffic patterns
Find ways to use bandwidth without causing problems
Avoid looping
Use all available paths<br>
slide28. Parallelism Within Networks: Where Is Bandwidth Needed? Depends upon usage pattern
If most activity is rack-local
At the rack switch
3-Tier might be fine
If most of the activity is east-west
Need more capacity at higher level(s), e.g. core and/or aggregation, etc
If most of activity is concentrating at egress (or ingress)
Need more core upstream capacity (supported by core)<br>
slide29. Parallelism Within Networks: How To Use Parallelism? Goals:
Must avoid cycles
Usage should be divided among paths
Ways to divide usage
Each use, e.g. network session, picks different path
Uses divide their traffic among parallel paths<br>
slide30. Using Parallelism Within Networks: Solvable at Different Levels (Examples) MPTCP – TCP that shares multiple paths for single session
Creates many sub-sessions and multiplexes overarching session among them
PORTland network – Link-Layer solutions
Path is based upon sender’s location in network<br>
slide31. Today Definitions
Data Center Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide32. Routing As we leave the domain of a single network, we enter the domain of internetworking
Many techniques for global connectivity are possible, but “packet switching” is the most common for Internet traffic as it crosses providers
Other types of switching, including (and perhaps especially) virtual circuits, e.g. ATM, are commonly used within provider networks.
Circuit switching is a fading technology, e.g. POTS, ISDN.
Packet switching involves a “router” making hop-by-hop decisions about how to “route” a packet.<br>
slide33. Routers May Also be called “Internet Switches” or “Network-layer switches”
May also perform other roles
Rule-based Firewalls
Rule-based load-balancing
Network Address Translation (NAT)
Virtual Private Networks and Tunneling<br>
slide34. Routing Protocols Routers exchange information with each other about the state of the network
Routers collect this into routing tables
Algorithms reduce the routing tables to simple forwarding tables that the switch can quickly consult w.r.t. which port should be the destination port for a particular incoming packet.<br>
slide35. Data Plane Vs Control Plane Routing algorithms run in the Control Plane
The control plane decides the rules for forwarding, but it does not do the forwarding
The Data Plan is where the forwarding actually occurs
Packets move in and out via the data plane, which consults the distilled tables provided by the control plane<br>
slide36. Today Definitions
Data Center Traffic Patterns
Data Center Fabric
Managing Parallelism Within Networks
Routing and Routers/Internet Switches (This is probably mostly review)
Looking forward to next class<br>
slide37. Next class Routers and switches need to sustain a very, very high throughput.
They need to very rapidly
Source a packet
Decide what to do with it
Sink it to another port
This could involve buffering it at the input port, for a decision, and/or at the output port
This could involve a computer-like shared memory system or a more purpose-built hardware switch<br>
slide38. Next class, cont. Next class, we are going to take a look a the architectures, where the points of contention are within them, discuss their performance and relative costs, &c.<br>