Professional Documents
Culture Documents
14-740: Networks: Lecture 24 Spring 2018 Kesden
14-740: Networks: Lecture 24 Spring 2018 Kesden
14-740: Networks: Lecture 24 Spring 2018 Kesden
Aggregation
Leaf
DC Topology: Venerable 3-Tier
• Can add a redundant core for increased throughput and resilience
Core
Aggregation
Leaf
DC Topology: Venerable 3-Tier
• 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
1x Switch Throughput
Wx Switch Throughput
W 2x Switch Throughput
Clos Networks
• 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
• Still an expensive expansion and not likely
to need all throughput capacity
By Piggly (talk) (Uploads) - Transferred from en.wikipedia to Commons.,
Public Domain, https://commons.wikimedia.org/w/index.php?curid=61536102
simultaneously
Leaf and Spine
Leaf and Spine
• Type of Clos network
• Essentially folded, but still N-to-N connections
• Derived from old phone company architecture, invented in 1950s.
• All paths are same length from edge to edge
• Great for switch vendors
• Need to pick path, as can choose any middle router
• Very redundant
• Can implement at layer-2 or layer-3
• More soon
Fat-Tree Networks
• More throughput at higher levels, more even across levels
• Not easy to do since buying more powerful switches is harder
• To extent possible, more cost per unit capacity
• Not possible beyond a modest point
• This is somewhat necessarily the case as, if bigger switches were more readily available
and economical, they’d be used at the bottom, and we’d be back where we started.
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
Note the replacement of
aggregation layer
(K/2)2 core routers Fat Tree (K=4) switches with 2 layers of
K/2 K-port switches
(K/2)2 servers
per pod
Position 0 1 0 1 0 1