Presenting: Ehud Shavit - PowerPoint PPT Presentation

1 / 43
About This Presentation
Title:

Presenting: Ehud Shavit

Description:

Tsutomu Yoshinaga - UEC - Tokyo. Todd C. Mowry - CMU. Vincent H. Berk - Dartmouth C. ... Perfect shuffle. s(x)=(an-1,an-2,..., a1,an) Inverse perfect shuffle ... – PowerPoint PPT presentation

Number of Views:29
Avg rating:3.0/5.0
Slides: 44
Provided by: techn7
Category:

less

Transcript and Presenter's Notes

Title: Presenting: Ehud Shavit


1
Interconnected NetworksTopologies
The Technion IIT Electrical Engineering
Department 048879 VLSI Architecture Seminar
  • Presenting Ehud Shavit

2
W.J. Dally
Main Reference
Elsevier, 2004
3
Acknowledgements
  • Tsutomu Yoshinaga - UEC - Tokyo
  • Todd C. Mowry - CMU
  • Vincent H. Berk - Dartmouth C.
  • Daniel J. Sorin - Duke
  • B. Ravikumar - Sonoma S.U.
  • William J. Dally - Stanford

4
Agenda
  • Introduction
  • General Properties
  • Traffic Patterns
  • Terminology
  • Performance
  • Common Topologies
  • Butterfly Topologies
  • Torus and Mesh Topologies

5
Agenda
  • Introduction
  • General Properties
  • Traffic Patterns
  • Terminology
  • Performance
  • Common Topologies
  • Butterfly Topologies
  • Torus and Mesh Topologies

6
Introduction (1)
  • Interconnected NetworkGenerally a Host Area
    Network
  • Connecting Processors and memory(and other
    devices)
  • May or may not share memory
  • Low Latency and High Bandwidth wanted
  • The well-known concept closest to NOC

7
Introduction (2)
  • Parallel Computing Requires FastCommunication
    Between Internal Nodes (Processors and Memories)
  • Direct Point-to-Point - O(n2) Wires
  • Connections between selected Pairs Routing
    through Intermediate Nodes
  • Network Topology Partly Determines Latency and
    Bandwidth

8
Introduction (3)
  • Applications
  • Supercomputers
  • Concurrent Computation
  • Future of computers inter-device communication

MIT J-Machine (1991)
9
Dally is Saying
  • We can build networks with 2-4 clocks/hop latency
    (12-24 clocks for a 512-node 3-cube)
  • networks faster than main memory access
  • need end-to-end hardware support to see this
  • Bandwidth of 4GB/s or more per channel (512GB/s
    bisection) is easy to achieve
  • nearly flat memory bandwidth
  • Topology is a matter of matching pin and
    bisection constraints to the packaging technology
  • its hard to beat a 3-D mesh or torus

10
Agenda
  • Introduction
  • General Properties
  • Traffic Patterns
  • Terminology
  • Performance
  • Common Topologies
  • Butterfly Topologies
  • Torus and Mesh Topologies

11
General Properties (1)
  • Generally Peer-to-Peer Communication
  • Single Owner
  • Normally 23 212 Nodes
  • Nodes are computer devices
  • High-end Nodes
  • Security - a non-issue

12
General Properties (2)
  • Generally known Traffic Patterns
  • Known and planned Topology
  • Every nodes task is normally known
  • Resources not low
  • Wired and Stationary
  • Packaging is a big issue

13
Agenda
  • Introduction
  • General Properties
  • Traffic Patterns
  • Terminology
  • Performance
  • Common Topologies
  • Butterfly Topologies
  • Torus and Mesh Topologies

14
Traffic Patterns (1)
15
Traffic Patterns (2) Shuffle
000
000
000
000
001
001
001
001
010
010
010
010
011
011
011
011
100
100
100
100
101
101
101
101
110
110
110
110
111
111
111
111
Perfect shuffle s(x)(an-1,an-2,, a1,an)
Inverse perfect shuffle s-1(x)(a1,an
an-1,an-2,, a2)
16
Traffic Patterns (3) Butterfly
000
000
000
000
001
001
001
001
010
010
010
010
011
011
011
011
100
100
100
100
101
101
101
101
110
110
110
110
111
111
111
111
butterfly (bit reversal) ß3(x)(a1 , a2 , a3)
?3(x)
2 sub-butterfly ß2(x)(a3, a1, a2)
17
Traffic Patterns (4) Shift
000
000
000
000
001
001
001
001
010
010
010
010
011
011
011
011
100
100
100
100
101
101
101
101
110
110
110
110
111
111
111
111
shift a(x) x1 (23)
2 sub-shift a 2(x) x1 (22)floor(x/ 22) 22
18
Traffic Patterns (5) Exchange
000
000
000
000
001
001
001
001
010
010
010
010
011
011
011
011
100
100
100
100
101
101
101
101
110
110
110
110
111
111
111
111
E3(x) (a3, a2, a1)
E2(x) (a3, a2, a1)
19
Agenda
  • Introduction
  • General Properties
  • Traffic Patterns
  • Terminology
  • Performance
  • Common Topologies
  • Butterfly Topologies
  • Torus and Mesh Topologies

20
Terminology (1)
  • Link cable with connectors on each end
  • Switch connect k inputs to k outputs
  • Phit Minimum of bits physically moved across
    link in one cycle
  • Packet Unit that requires routing information,
    some number of Phits
  • Topology The mathematical structure of the
    network

21
Terminology (2)
  • Frequency The rate at which bits are transported
    thru channel c.
  • Cut A set of channels, if removed, dividing
    network into two disjoint parts
  • Bisection min. number of links that, if removed,
    would separate the network
  • Direct All switches attached to host nodes
  • Indirect Many switches not attached to host nodes

22
Terminology (3)
  • Degree outgoing links
  • Diameter number of links crossed between nodes
    on maximum shortest path
  • Average Hop-Count number of hops to random
    destination
  • Critical Length max length of a wire for a
    given nominal frequency

23
Terminology (4)
  • Channel Load The ratio between the demanded and
    maximum bandwidth
  • Channel Width The number of signal running thru
    a channel
  • Ideal Throughput the input bandwidth that just
    saturates the bottleneck channel

24
Terminology (5)
  • Latency The time required for a bit to travel
    thru channel c.
  • Average latency without load
  • Routers Latency
  • Time of Flight
  • Serialization Latency

25
Terminology (6) - Latency
26
Agenda
  • Introduction
  • General Properties
  • Traffic Patterns
  • Terminology
  • Performance
  • Common Topologies
  • Butterfly Topologies
  • Torus and Mesh Topologies

27
Performance
  • Degree
  • Bisection
  • Average Path Hops Count
  • Channel Load
  • Channel Width
  • Ideal Throughput
  • Average Latency

28
Agenda
  • Introduction
  • General Properties
  • Traffic Patterns
  • Terminology
  • Performance
  • Common Topologies
  • Butterfly Topologies
  • Torus and Mesh Topologies

29
Common Topologies (1)
30
Common Topologies (2)
31
Agenda
  • Introduction
  • General Properties
  • Traffic Patterns
  • Terminology
  • Performance
  • Common Topologies
  • Butterfly Topologies
  • Torus and Mesh Topologies

32
Butterfly Topology (1)
2-ary 2-fly
33
Butterfly Topology (2)
  • k-ary n-flies Butterfly Network
  • An n stage of radix k switches butterfly network.
  • Such a network is composed of Nkn
    source/destination terminal nodes
  • n stages of kn-1 switching nodes each with k
    inputs and k outputs
  • All (n1)N channels are unidirectional.

34
Butterfly Topology (3)
2-ary 4-fly
35
Butterfly Topology (4)
  • All paths equal length
  • Unique path from any input to any output
  • Conflicts cause tree saturation
  • This may be solved using extra stages
  • Different forms to do that (After, between)

36
Agenda
  • Introduction
  • General Properties
  • Traffic Patterns
  • Terminology
  • Performance
  • Common Topologies
  • Butterfly Topologies
  • Torus and Mesh Topologies

37
Torus/Mesh Topo. (1)
A Simple 8 node ring
38
Torus/Mesh Topo. (2)
  • k-ary n-cube Torus Network
  • An n dimensional of radix k torus network.
  • Such a network is composed of Nkn nodes
  • n dimensional rings-structure
  • k nodes along each ring.
  • Each node both a terminal and a switch.
  • Bidirectional 2nN channels.

39
Torus topology
Torus/Mesh Topo. (3)
2D (4-ary 2-cube)
3D (3-ary 3-cube)
40
Torus/Mesh Topo. (4)
  • k-ary n-mesh Mesh Networ
  • Like a torus but with no ring closing.
  • Number of channels

41
Mesh topology
Torus/Mesh Topo. (2)
node
2D
3D
42
Conclusions
  • Interconnected Networks and NOC are close
    relatives
  • We should learn and try to duplicate
  • Known Topologies
  • Known Traffic Patterns
  • Known Routing Protocols
  • etc...

43
Networks MeshButterflies
KarlFriedrich Hieronymus Baron of Munchausen
(1720-1797)
Write a Comment
User Comments (0)
About PowerShow.com