Lecture 6: Vector

Download Report

Transcript Lecture 6: Vector

Lecture 7:
Interconnection Network
Part I: Basic Definitions
Part II: Message Passing Multicomputers
1
Part I: Basic Definitions
A network is characterized by its
topology, routing algorithm,
switching strategy, and flow control
mechanism.
2
Basic Definitions

Topology: the physical interconnection structure of
the network graph; regular (most parallel machines)or
irregular (WAN).

Routing algorithm: determines which routes a
message may follow through the network graph.

Switch strategy: determines how the data in a
message traverses its route (circuit or packet switch)

Flow control: determines when the message, or
portion of it, move along its route.
3
Various Interconnect
Topologies
N/2
Butterfly
°
°
°
N/2
Butterfly
°
°
°
4
Routing Messages
 Shared
Media
– Broadcast to everyone
 Switched
media : Needs real routing.
– Circuit switching
– Packet switching
5
System Architecture
System Bus
processor
Memory
Bus
Adapter
SCSI
controller
I/O Bus
Network
Interface
Card
Transmission Media
Network
Switch
6
Shared-Media Networks
 Allow
single message transmission at a time
 Bus-based
networks: Ethernet, Fast Ethernet
 Ring-based
networks: IBM Token ring, FDDI
7
Shared-Media Network
 Advantage
– simple design
– less expensive
– simple routing
– nature for broadcast
and multicast
– scalable, but limited,
within each segment
 Disadvantages
– fixed channel
bandwidth
– need router or
gateway to go beyond
each segment
– limited distance span
8
Switched Network
 Allow
simultaneous transmission of many
messages
 Typical
switch size: 8 to 64 ports
9
Types of Switches
 Cell-based
switching
– fixed-size packets
– e.g., ATM switches (53-byte cells)
 Frame-based
(Packet-based) switching
– variable-size packets
– e.g., Switched Ethernet (e.g., HP EtherTwist)
– e.g., FDDI switch (e.g., DEC GIGAswitch)
– e.g., Myrinet switch (wormhole routing)
10
Generic Switch Architecture
Input
buffer
4-port
switch
Output
buffer
Logical
Crossbar
Organization
Control Unit
11
Buffer Architecture
 Input
buffer:
–natural design: FIFO
–random access buffer (more expensive)
 Output
buffer: more complicated
–need better performance
–solve output contention
12
Buffer Architecture
 Dedicated
buffer (for each port)
–ease of routing
–guaranteed service per channel
 Shared
buffer
–better buffer utilization
–one channel burst may take all buffer
13
Logical Crossbar (ATM 16-port)
155 Mbps
2.5 Gbps Bus
622 Mbps
Time Division Bus
14
Other Switch Design
 Shared-memory
Switch
– CNET Prelude switch
 2-D
mesh crossbar
– Myrinet, DEC GIGAswitch
 Clos
network (scalable)
– multistage networks
 Bene
Network:
– Washington Univ. Gigabit ATM Network
15
Cautions on Speeds
 The
actual application-level data rate is less
than advertised speed
– ATM: 155 Mbps ==> at most 134 Mbps (14%)
 Switch
delay: from input port to output port
– Myrinet: 100 nsec (8-port)
– ATM: 10-30 microseconds
– WU Gigabit ATM: 10-20 microseconds
16
Packet Routing in Switched
Network
17
Circuit switching
 Set
up; communication; release
 Circuits reserved for communication
 Advantages:
short delays (after set up)
 Disadvantage: not efficient for bursty
traffic due to the long setup time.
18
Packet Switching (Datagram)
 Put
addresses in packets; route one by one
 Switch determines the path
 Deterministic:
always follow same path
 Adaptive: pick different paths to avoid congestion,
failures
 Randomized routing: pick between several good
paths to balance network load


Adv: efficient; robust against failure
Disadv: delay variations; misordering possible.
19
Deterministic
 Circuit
established from source to destination,
message picks the circuit to follow
– Determined based on source and destination address
– All packets follow the same route.
 Adv:
efficient; ordered; smaller jitter
 Disadv:
setup time; not robust; scalability.
20
Deterministic Routing
Examples
 Mesh:
– dimension-order routing
– (x1, y1) -> (x2, y2)
– first x = x2 - x1,
– then y = y2 - y1,
 Hypercube:
–
–
–
–
edge-cube routing
X = xox1x2 . . .xn -> Y = yoy1y2 . . .yn
R = X xor Y
Traverse dimensions of differing
address in order
 Tree: common ancestor
110
010
111
011
100
000
001
101
21
Store and Forward vs. CutThrough
 Store-and-forward:
– each switch waits for the full packet to arrive in switch
before sending to the next switch (good for WAN)
 Cut-through:
– switch examines the header, decides where to send
the message, and then starts forwarding it immediately
– Two approaches: (1) Virtual cut-through (2)
Wormhole routing
flit
flit
H
D
flit
D
D CRC
Packet
22
Store and Forward vs. CutThrough
Node 4
(destination)
H
H
Node 3
Node 2
Node 1
(source)
H
H
a
b
Node 4
(destination)
Node 1
(source)
H
b
b
b
c
c
c
(a) Store-and-Forward
c
H
a
b
H
a
b
c
H
a
b
c
a
b
c
Node 3
Node 2
a
a
a
c
H: header
a, b, c : data elements
(b) Cut-through (wormhole)
23
Switching Mechanisms
 Store-and-forward
 Cut-through
– buffer each packet
– small buffer
– buffer management
– low latency
– support link-level ack
– no link-level ack
– good for networks
with high error rate
(e.g., WANs)
– good for networks
with very low error
rate (e.g., LANs)
24
(1) Virtual Cut-through
 To
spool the blocked incoming packet into input
buffer
 The behavior under contention degrades to that of
store-and-forward.
 Requires a buffer large enough to hold the largest
packet.
25
(2) Worm-hole Routing
 The
packet s subdivided into smaller flits.
 The
header flit knows where the train (packet) is
going. All the data flits follow the header flit.
 Different
packets can be interleaved but flits from
different packets cannot be mixed up.
 When
head of message is blocked, it leaves the
tail of the message in-place along the route.
 Potentially
blocking other messages
 Needs
only buffer the piece of the packet that is
sent between switches.
26
Performance Comparison

Let
– L= packet length
– W= channel bandwidth
– D= distance (no. of nodes -1)




T store&forward= (L/W)(D+1)
Twormhole = (L/W) + (F/W)xD
If L>>F; Twormhole = (L/W)
Implication: distance insensitive
27
Store and Forward vs. CutThrough
 Advantage
–Latency reduces from function of:
number of intermediate switches X by the size of the packet
to
time for 1st part of the packet to negotiate the switches
+ the packet size ÷ interconnect BW
28
Congestion Control
 Packet
switched networks do not reserve
bandwidth; this leads to contention
 Solutions:
– Packet discarding: If packet arrives at switch
and no room in buffer, packet is discarded (e.g.,
UDP)
– Flow control: prevent packets from entering
until contention is reduced (e.g., freeway onramp metering lights)
29
Flow control:
 Between
pairs of receivers and senders; use
feedback to tell sender when allowed to send next
packet
 Back-pressure: separate wires to tell to stop
 Window: give original sender right to send N packets
before getting permission to send more.(TCP)
 Choke packets: Each packet received by busy switch
in warning state sent back to the source via choke
packet. Source reduces traffic to that destination by a
fixed % (e.g., ATM)
30