LDM Link Division Multiplexing

1
LDM Link Division Multiplexing
MATRICS Research Group
Technion Israel Institute of Technology

• Arkadiy Morgenshtein, Avinoam Kolodny, Ran Ginosar

MATRICS Research Group, Electrical Engineering
Department Technion Israel Institute of
Technology Haifa, Israel
2
Background Motivation
3
Networks-on-Chip (NoC)

• NoC Characteristics
• Packets-based data routing
• Modules connected by routers network
• Shared links
• Supports QoS communication - QNoC

4
Quality of Service in NoC

• QNoC NoC with QoS
• Signaling urgent short packets
• Read-Time audio/video applications
• Read/Write memory and register access
• Block-Transfer long blocks of data

low latency, high priority
latency ?, priority ?
latency ? ?, priority ? ?
high latency, low priority
5
Motivation
Data flow in QNoC links

• Data is transmitted using Time-Sharing

• At each time slot all the wires are dedicated to
  a single source

• QoS priority defines order and duration for each
  source

- Low link utilization
Link Division Multiplexing (LDM)
solution
- Timing dependency
- High power
6
LDM Concept
7
Link Division Multiplexing (LDM)

• Link resources (wires) are divided among QoS
  levels Link Division Multiplexing
• The link is composed of outputs of several
  serializers
• Each serializer is dedicated to transmission of
  data at certain QoS level
• The number of wires at each level is allocated
  according to QoS level priority.
• LDM allows simultaneous transport of data in
  various QoS levels.

8
Link Division Multiplexing (LDM)
m
X m-to-n serializers

• LDM allows dynamic division of resources
  according to QoS levels
• Full utilization of the link resources
• No timing dependency of lower QoS levels on
  higher levels
• Simultaneous data transport at different QoS
  levels while maintaining the throughput and
  latency demands
• Higher data transmission rate with improved
  efficiency of power control

9
LDM Architecture
10
LDM Transceiver
11
Controller Implementation
Trade-offs
Alternative
Reduced wiring overhead - Reduced data efficiency rate Add the control data to the packet. Each wire carries the information about the packet to which it is designated.
Reduced hardware overhead - Reduced flexibility and utilization Predefined allocation patterns of wires. Wires allocated according to operation mode without control communication.
High data efficiency rate flexibility - Increased wiring overhead Send control signals at dedicated wires. Additional wires used for control communication in the transceiver.
our architecture
12
QNoC Router with LDM Link
TDM
LDM

• TDM data is classified and stored in dedicated
  buffers according to QoS levels
• LDM various QoS levels are treated
  simultaneously, no need for separate storage
• Fewer buffers are needed in LDM

13
Results
14
Simulation Setup

• LDM communication environment was implemented
  and emulated in Matlab
• QNoC link with 32 wires connected between two
  routers with four clients each
• Two possible patterns of wires allocation were
  set
• 8,8,8,8
• 16,8,4,4
• Four QoS levels were used Signaling,
  Real-Time, R/W, Block-Transfer
• Parameters were defined for each QoS level
• Size of packet (in 32-bit flits)
• Probability of data appearance at given QoS
  level (including no data)
• Delay expressing the processing time of packet
  before transmission
• For each client a profile was built basing on
  set of five data probabilities

15
LDM in Various Data Scenarios
Flits transmitted Flits transmitted QoS probability QoS probability QoS probability QoS probability QoS probability Client scenario distribution type
TDM LDM Pblock-trans Pread/write Preal-time Psignaling Pno-data Client scenario distribution type
99876 99806 0.15 0.25 0.05 0.25 0.30 all A homogeneous
89963 89983 0.002 0.001 0.001 0.001 0.995 all B homogeneous
39792 55890 0.4 0.2 0.2 0.1 0.1 C1 C heterogeneous
39792 55890 0 0.005 0.001 0 0.994 C2 C heterogeneous
39792 55890 0 0.005 0.001 0 0.994 C3 C heterogeneous
39792 55890 0 0.005 0.001 0 0.994 C4 C heterogeneous
87572 89983 0.001 0.004 0.001 0.001 0.993 C1 D heterogeneous
87572 89983 0.001 0.001 0.002 0.006 0.99 C2 D heterogeneous
87572 89983 0.3 0.05 0.05 0.3 0.3 C3 D heterogeneous
87572 89983 0.001 0.004 0.004 0.001 0.99 C4 D heterogeneous

• Simulations contained data generation and
  transport during 100,000 clock cycles
• The simulation scenarios were divided into two
  types
• Homogeneous same QoS probability profiles for
  all clients
• Heterogeneous different QoS probability
  profiles for all clients
• Number of transported flits in LDM was increased
  by up to 40

16
Performance vs. Packet Delays
TDM
LDM

• LDM effectiveness was evaluated as function of
  packets delay before transmission.
• Number of transported flits in LDM was increased
  by up to 50
• LDM link has maximum value for certain delay.
• for low delays there is a queue of data in the
  buffer
• for higher delays the number of the flits
  reduces similarly to TDM.

17
Power Reduction in LDM
LDM with lower VDD
TDM
LDM
X m-to-n serializers
X m-to-n serializers
or
TLDM
TLDM_low
TTDM
sleep

• High link utilization in LDM
• Longer sleep mode allowed in LDM clock and
  supply gating
• Timing can be traded for Voltage Scaling to
  reduce power

TTDM TLDM Tsleep VTDM VLDM
TTDM TLDM_low VTDM gt VLDM_low
PLDM lt PTDM
18
Summary

• Link Division Multiplexing proposed
• LDM targets improvement of link utilization
• Increase in data rate and reduction of power
• LDM link was implemented and simulated
• Number of transmitted flits increased by up to
  50

19
Questions?
m-to-1
n
X m-to-n serializers
m