Low-Latency Virtual-Channel Routers for On-Chip Networks
Robert Mullins, Andrew West, Simon Moore Presented by
Sailesh Kumar
Outline
s
Motivation
Why Network-on-chip (NoC)
Comparison to Packet Networks
Similarities Differences Design Constraints
Topology and Routing/Switching techniques for NoC
Mesh, fat-tree, honey-comb Greedy, Deflection, Wormhole, Virtual-Channels
Start with the paper Design of a Low-Latency Virtual-Channel Router
2 - Sailesh Kumar - 02/28/13
Why NoC
s
Billion transistor era has arrived
Several such SoC are in pipeline, Inter-connection is critical
A generic inter-connection architecture ensures
Reduced design time IP reuse Predictable backend (versus ad-hoc wiring)
Bus based inter-connects were sufficient until now But not now
Shared bus is slow (arbitrates between several requesters) More components increase loading => speed drops further Ad-hoc routing of wires results in backend complications, lower performance and higher power consumption
3 - Sailesh Kumar - 02/28/13
Why NoC (cont)
s
Recently Dally proposed an idea
Route packets not wires as in data networks
Point to point communication
Point to point links are faster
s
Create a chip wide network (Like a regular IP WAN)
A router at every node Links connecting all routers Messages encapsulated in packets, which are routed
Challenges
Topologies, Routing protocol Network and router design with small footprint and low latency
4 - Sailesh Kumar - 02/28/13
Some more motivations
s
The need to put repeaters into long wires allows us to add the switching needed to implement a network at little additional cost Makes efficient use of critical global wiring resources by sharing them across different senders and receivers Simplifies overall design Design a single router and do copy-paste in both dimension
s s
5 - Sailesh Kumar - 02/28/13
A typical NoC node
s
Layered Design of reconfigurable micronetworks. Exploits methods and tools used for general network. Micronetworks based on the ISO/OSI model. NoC architecture consists of Physical, Data link, and Network layers.
6 - Sailesh Kumar - 02/28/13
A typical NoC
s
Layered Design of reconfigurable micronetworks. Exploits methods and Implemented in tools used for general cores, enables endto-end reliable network. transport Micronetworks based on the ISO/OSI model. NoC architecture consists of Physical, Data link, and Network layers.
7 - Sailesh Kumar - 02/28/13
A typical NoC
s
Layered Design of reconfigurable micronetworks. Exploits methods and Implemented in tools used for general cores, enables endto-end reliable network. transport Micronetworks based on the ISO/OSI model. NoC architecture of Physical, Data and Network layers.
Multi hop route consists packet setup, link, addressing, etc
8 - Sailesh Kumar - 02/28/13
A typical NoC
s
Layered Design of reconfigurable micronetworks. Exploits methods and Implemented in tools used for general cores, enables endto-end reliable network. transport Micronetworks based on the ISO/OSI model. NoC architecture of Physical, Data and Network layers.
Multi hop route consists packet setup, link, addressing, etc Contention issues, reliability issues, grouping of physical layer bits, e.g. flits
9 - Sailesh Kumar - 02/28/13
A NoC topology
s
s s
Cores Communicates With Each Other Using NoC NoC Consists of Routers (R) and Network Interfaces (NI) A NI linked to Router by Non-Pipelined Wires One or More Cores Connected to a NI
10 - Sailesh Kumar - 02/28/13
Another NoC topologies
Fat tree
Mesh
Multi hop route setup, packet addressing, etc
11 - Sailesh Kumar - 02/28/13
Routing protocols
s s
We will only consider mesh topology Objective is to find a path from a source to a destination Greedy Algorithms (deterministic)
Choose shortest path (e.g. X-Y)
s s
Adaptive routing
If congestion, choose alternative path Deflection routing Is adaptive better than greedy => NOT REALLY (when only local information is used)
Adaptive routing can also result in livelock
12 - Sailesh Kumar - 02/28/13
Switching techniques
s
Circuit Switching: A control message is sent from source to
destination and a path is reserved. Communication starts. The path is released when communication is complete.
Store-and-forward policy (Packet Switching): each switch waits for the full packet to arrive in switch before sending to the next switch Cut-through routing or worm hole routing: switch examines the header, decides where to send the message, and then starts forwarding it immediately
In worm hole routing, when head of message is blocked, message stays strung out over the network, potentially blocking other messages (Needs only buffer the piece of the packet that is sent between switches). Cut through routing lets the tail continue when head is blocked, storing the whole message into an intermediate switch. (Need buffer large enough to hold the largest packet).
13 - Sailesh Kumar - 02/28/13
Wormhole Routing Good fit for NoC
s
Wormhole routing is good for NoC
Low latency Less buffering requirements
Suffers from deadlock
14 - Sailesh Kumar - 02/28/13
[example from Li and McKinley, IEEE Computer v26n2, 1993]
Adding Virtual Channels
s s
With virtual channels, deadlock can be avoided Move message and reply on different channels => Will never have loop on a single channel
Packet switches from lo to hi channel
15 - Sailesh Kumar - 02/28/13
Designing Virtual Channel Routers
s
Design Constraints in NoC
Minimize Latency Minimize Buffering Minimal footprint
s s s
Can exploit far greater number of pins and wires May use fat data and flow control wires Objective: Design routers with minimal latency
This will also result in smaller buffers
This paper presents design of a low latency router
Cycle time of 12 FO4 Single cycle routing/switching
16 - Sailesh Kumar - 02/28/13
Designing Virtual Channel Routers
s
Design Constraints in NoC
Minimize Latency Minimize Buffering Minimal footprint
s s s
Can exploit far greater number of pins and wires May use fat data and flow control wires Objective: Design routers with minimal latency
This will also result in smaller buffers
This paper presents design of a low latency router
Cycle time of 12 FO4 Single cycle routing/switching
17 - Sailesh Kumar - 02/28/13
A Virtual Channel Router
18 - Sailesh Kumar - 02/28/13
Designing Virtual Channel Routers
Every VC of every input port has buffers to hold arriving flits
Arriving flits are placed into the buffers of corresponding VC
19 - Sailesh Kumar - 02/28/13
Designing Virtual Channel Routers
Every VC of every input port has buffers to hold arriving flits Routing logic assigns set of outgoing VC on which flit can go Arbitrates between competing input VC & allocates output VC Arriving flits are placed into the buffers of corresponding VC
20 - Sailesh Kumar - 02/28/13
Designing Virtual Channel Routers
Every VC of every input port has buffers to hold arriving flits Routing logic assigns set of outgoing VC on which flit can go Arbitrates between competing input VC & allocates output VC Arriving flits are placed into the buffers of corresponding VC Matches successful input ports (allocated VC) to output ports
Flits at input VCs getting grants are passed to output VCs
21 - Sailesh Kumar - 02/28/13
Routing Logic
s
Three possibilities
Return a single VC Return set of VCs on a single port Return any VCs
Look ahead routing
Routing performed at the previous router Good for X-Y deterministic (non adaptive) routing A SGI routing chip first implemented it
22 - Sailesh Kumar - 02/28/13
VC Allocation
s s
Complexity of VC allocation depends on routing range Routing returns single VC
Needs PxV input arbiter for every outgoing VC
Routing returns multiple VC at single port
Additional V:1 arbiter at every input VC to reduce potential outgoing VC to 1
Routing returns any set of VCs
Needs two cascaded PxV input arbiters
We consider multiple VC at single port case
23 - Sailesh Kumar - 02/28/13
VC Allocation Logic
At every outgoing VC following logic is needed
24 - Sailesh Kumar - 02/28/13
Switch Allocation
s
Individual flits at input VCs arbitrate for access to the crossbar port Arbitration can be performed in two stages First stage
A VC among V possible VCs at every input port is selected V:1 arbiter at every input port
s s
Second stage
Winning VC at every input port is matched to the output port P:1 arbiter at every output port
s s
This scheme doesnt guarantee a maximal/maximum/good matching But simple to implement
25 - Sailesh Kumar - 02/28/13
Switch Allocation
26 - Sailesh Kumar - 02/28/13
Issues
s s s s
VC allocation and Switch allocation are serialized A flit will either take 2 clocks to get through Else clock speed will be low Solution: Speculative switch allocation
27 - Sailesh Kumar - 02/28/13
Speculative Switch Allocation
s
Dally proposed speculative switch allocation
Perform switch and VC allocation in parallel Assume that participating VC in switch allocation will get the output VC If not then wasted cycle
An even better idea is to perform speculative and non-speculative switch allocation in parallel
Non-speculative allocation has higher priority Note that non-speculative allocation is done for input VCs which has already been allocated an output VC Speculative
will work
s s
Mostly one cycle delay under light load Mostly one cycle delay under heavy load
Non-speculative will work
28 - Sailesh Kumar - 02/28/13
s s
Further Enhancement Is it possible to have zero cycle VC/switch allocation YES, Most of the time, thats what this paper is about!
29 - Sailesh Kumar - 02/28/13
Idea 1: Free Virtual Channel Queue
s
Keep queue of free VC at every outgoing port
Also bit mask with one set bit
Thus First stage of VC allocation where an output VC is selected will be removed
30 - Sailesh Kumar - 02/28/13
Idea 1: Free Virtual Channel Queue
s
Keep queue of free VC at every outgoing port
Also bit mask with one set bit
Thus First stage of VC allocation where an output VC is selected will be removed
31 - Sailesh Kumar - 02/28/13
Idea 2: Pre-computing arbitration decisions
s
If somehow, you know the arbitration results before flits actually arrive and fight for the VC and switch I mean, every arbitration decision
VC allocation Switch allocation Etc
Then the router can be made to run in zero cycle
The arriving flit route/switch in the same clock they arrive
Also, clock speed may be pretty good
Data path and control path are no more in series
Thats what the idea 2 is.
32 - Sailesh Kumar - 02/28/13
Some preliminaries before going into detail
Tree Arbiters
Implements large arbiters using tree of small arbiters
Matrix Arbiters
Fair and Fast arbiter implementation
33 - Sailesh Kumar - 02/28/13
VC allocation using a Tree Arbiter
34 - Sailesh Kumar - 02/28/13
A Matrix Arbiter
35 - Sailesh Kumar - 02/28/13
Pre-computing arbitration decisions
s
An alternative arbiter design
36 - Sailesh Kumar - 02/28/13
Pre-computing arbitration decisions
s
An alternative arbiter design
Grants are product of latched enables and the requests
Generate grant enables one cycle prior and latch them
37 - Sailesh Kumar - 02/28/13
Pre-computing arbitration decisions
s
An alternative arbiter design
Grants are product of latched enables and the requests
Generate grant enables one cycle prior and latch them
Grants are generated in same clock as request arrives
If at least one request remains
38 - Sailesh Kumar - 02/28/13
Pre-computing arbitration decisions
s
An alternative arbiter design
Grants are product of latched enables and the requests
Generate grant enables one cycle prior and latch them
However, when no request remains, it is difficult to generate grant enables ???
39 - Sailesh Kumar - 02/28/13
Generating grant enables
s
Safe Environment
Only one request may arrive in a cycle Thus it is safe to assert all grant enables Thus grant can still be generated in same cycle
Unsafe Environment
Multiple request may arrive in same cycle Can still assert all grants But need to abort when multiple requests arrive in same cycle
s s
All first stage V:1 arbiters operate under safe environment However P:1 arbiters doesnt
40 - Sailesh Kumar - 02/28/13
Generating grant enables
s
Even in unsafe environments, assert all grants
May need to abort when multiple requests arrive Note that after an abort, a correct arbitration is ensured in the next cycle
s s s
Why will it work? Because in lightly loaded network, multiple requests for same VC/port will not arrive (few aborts) In heavily loaded network flits will remain buffered and Non-speculative arbitration (higher priority) will happen most of the time
Few aborts again
41 - Sailesh Kumar - 02/28/13
I will skip the design details now
s s
Since it is confusing and complex Will jump to critical path analysis
42 - Sailesh Kumar - 02/28/13
Analysis of critical path
Generates VC/switch grants from precomputed grant enables
43 - Sailesh Kumar - 02/28/13
Analysis of critical path
Generates VC/switch grants from precomputed grant enables
Crossbar traversal is aborted once invalid grants are detected
44 - Sailesh Kumar - 02/28/13
Analysis of critical path
Generates VC/switch grants from precomputed grant enables In case of an abort, the correct control signals are ensured in the next cycle
Crossbar traversal is aborted once invalid grants are detected
45 - Sailesh Kumar - 02/28/13
Final design
s
Control path critical delay is 12 FO4
Until now, the best design had 20 FO4 delays
They have sampled a NoC based ASIC last week using this idea Runs at several GHz speeds Note that fast cycle time is possible by
Running VC allocation and Switch allocation in parallel Must use speculation, else delay will be higher (1 more cycle)
s s
46 - Sailesh Kumar - 02/28/13
Simulation results
47 - Sailesh Kumar - 02/28/13
If (doubts) Then Ask; Else Thank you; Goto Discussion; End if;
48 - Sailesh Kumar - 02/28/13