Implementation Analysis of NoC: A MPSoC TraceDriven Approach

1
Implementation Analysis of NoC A MPSoC
TraceDriven Approach

• Sergio Tota¹, Mario R. Casu¹, Luca Macchiarulo²

¹ Politecnico di Torino
² University of Hawaii
2
Outline

• Motivations
• Network-on-Chip Paradigm
• Definitions and Terminology
• NoC Topologies and Routing Strategies
• Switch Design
• Trace-Based Emulation Experiments
• Results
• Conclusions

3
Motivations

• The number of processors in the same die
  increases at each technology node (MPSoC)
• This trend leads to the need of a scalable
  communication infrastructure
• Shared bus are not a long-term solution
• On-Chip Micronetworks better suit the demand of
  scalability and performance

4
Network-on-Chip (NoC)

• On-chip networks inherit some of the features of
  computer networks
• New constraints emerge for the on-chip
  implementation (i.e. area, power)
• The NoC characteristics depend on the choice of
  the Topologies and Routing Strategy
• Regular On-chip networks facilitate modular
  design and improve performance

5
Definitions and Terminology

• Flit The elementary unit of information
  excanged in the communication network in a clock
  cycle.
• Packet An element of information that a
  processing element (PE) sends to another PE. A
  packet may consist of a variable number of
  flits.
• Switch The component of the network that is
  in charge of flit routing.

6
Definitions and Terminology (cont'd)

• Flit Latency The time needed for a FLIT to
  reach its target PE from its source PE.
• Packet Latency The time needed for a PACKET to
  reach its target PE from its source PE.
• Packet Spread The time from the reception of
  the first flit of a packet to the reception of
  the last one.

7
Network Topology
1
2
3
4
1
2
3
4
5
6
7
8
5
6
7
8
9
10
11
12
9
10
11
12
13
14
15
16
13
14
15
16
Mesh
Physical implementation
8
Network Topology (cont'd)
1
2
4
3
1
2
3
4
5
6
7
8
13
14
16
15
9
10
11
12
5
6
8
7
13
14
15
16
1
2
3
4
Torus
Physical implementation
9
Routing

• What is important to us
• Minimum latency is of paramount importance in
  MP-SoC (interprocess communication).
• Ideally 1 clock latency per switch (flit enters
  at time t and exits at t1)
• Maximum switch clock frequency (technology
  routing logic limits)
• Deadlock free
• No flits are ever lost once a flit is injected
  in the NoC, it will eventually reach its
  destination

10
Routing Strategy Wormhole

• In wormhole routing a header flit digs the
  hole
• Successive flits are routed to the same
  direction
• In case of blocks and lossless NoC we need
• Buffers
• A backpressure mechanism (unless you dont have
  infinite FIFOs)
• We assume X-Y static routing (first X then Y,
  proven deadlock free)

11
Worm-Hole
Src
Dest
12
Worm-Hole
Src
HF
F2
F3
F4
TF
Dest
13
Worm-Hole
Src
F2
HF
F3
F4
TF
Dest
14
Worm-Hole
Src
F3
F2
F4
TF
HF
Dest
15
Worm-Hole
Src
F4
F3
TF
F2
HF
Dest
16
Worm-Hole
Src
F4
F3
TF
F2
HF
Dest
17
Worm-Hole
Src
F3
F4
TF
F2
HF
Dest
18
Worm-Hole
Src
F4
TF
F3
F2
Dest
HF
19
Worm-Hole
Src
TF
F4
F3
Dest
F2
HF
20
Worm-Hole
Src
TF
F4
Dest
F3
F2
HF
21
Worm-Hole
Src
TF
Dest
F3
F2
HF
22
Worm-Hole
Src
Dest
TF
F3
F2
HF
23
Routing Strategy Deflection Routing

• Every flit can be routed to different directions
  (no packet notion at the switch level)
• if the optimal direction is blocked, the flit is
  deflected to another direction
• switch latency of 1 clock cycle no matter the
  congestion
• minimum buffer requirements
• A.K.A. Hot Potato, deadlock free by
  construction

24
Hot-Potato
Src
Dest
25
Hot-Potato
Src
HF
F2
F3
TF
Dest
26
Hot-Potato
Src
F2
HF
F3
TF
Dest
27
Hot-Potato
Src
F3
F2
HF
TF
Dest
28
Hot-Potato
Src
TF
HF
F2
F3
Dest
29
Hot-Potato
Src
TF
HF
F2
F3
Dest
30
Hot-Potato
Src
TF
Dest
HF
F2
F3
31
Hot-Potato
Src
TF
Dest
HF
F2
F3
32
Hot-Potato
Src
Dest
TF
HF
F2
F3
33
Hot-Potato
Src
Dest
F3
TF
HF
F2
34
Routing Techniques
Wormhole
Hot-Potato
No packets reordering - Static routing -
Buffering ( ?2 flits/port) - Back pressure - XY
routing needs mesh
- Packets reordering Adaptive routing No
buffering No back pressure Works with
torus/mesh
35
Switch logic scheme
36
Physical Implementation(IBM CMOS 0.13 ?m _at_ 500
MHz) 160 Gbit/s
Deflection-Routing
Wormhole10²
Wormhole 2¹
0.038 mm²
0.273 mm²
0.086 mm²
Area _at_ 64 bits
0.07 mm²
0.502 mm²
0.140 mm²
Area _at_ 128 bits
0.14 mm²
0.910 mm²
0.234 mm²
Area _at_ 256 bits
29 uW/MHz
190 uW/MHz
54 uW/MHz
Power _at_ 64 bits
52 uW/MHz
380 uW/MHz
92 uW/MHz
Power _at_ 128 bits
102 uW/MHz
655 uW/MHz
171 uW/MHz
Power _at_ 256 bits
¹Two buffers per port
²Ten buffers per port
37
Emulation environment
38
Emulation Environment

• Statistically generated traffic is not realistic
• Standford SPLASH-2 MultiProcessor Benchmark
  (radix,lu,fft,ocean,raytrace)
• RSIM cycle-accurate simulator (UIUC) to extract
  traffic traces between processors
• Simulation in VHDL using behavioural traffic
  generators and RTL NoC implementations
• 1 dual Opteron and 5 dual Xeon servers with Linux
  64 bit, Modelsim, Synopsys and Encounter

39
Experiments parameters

• NoC size N x N, N2 and N4
• Packet size 10 flits
• Wormhole buffer size 2 flits/port (minimum
  allowed value)
• Standard simulation PE computation time 1 X
• Accelerated simulation PE computation time 20
  X, same NoC clock frequency (worse traffic
  conditions)
• WH wormhole, HP hot potato

40
Results
Mesh uniform traffic ideal latency
Torus uniform traffic ideal latency
Mean Lat. 2/3N Peak Lat. 2N-2
Mean Lat. 1/2N Peak Lat. N
41
Results
Ideal packet latency packet size 10
42
Comments

• Peak flit latency occurs sporadically (peak gtgt
  average)
• WH and HP show comparable latency (both peak and
  ave)
• HP packet mean latency ? 10 flits arrive
  in-order (with few exceptions)
• No substantial differences between standard and
  accelerated simulations
• Benchmark execution time overhead due to NoC
  congestion negligible
• Overall, WH is not better nor worse than HP

43
Conclusions

• RTL NoC based on Worm-Hole and Hot Potato
• Switch design and synthesis on 0.13 ?m CMOS
• 500 MHz clock frequency
• Real MP-SoC trace simulations
• WH and HP show similar performance but HP needs
  less area and power
• The strength of HP is supposed to emerge in a
  condition of higher load Need for benchmarks
  that generate higher traffic
• Ongoing work
• New RTL design working at 700 MHz
• Reconstruction interface for HP implemented