1. Low-Latency Virtual-Channel Routers for On-Chip Networks Robert Mullins, Andrew West, Simon Moore Presented by Sailesh Kumar

2. 2 - Sailesh Kumar - 8/8/2015 Outline n Motivation »Why Network-on-chip (NoC) n Comparison to Packet Networks »Similarities »Differences »Design Constraints n Topology and Routing/Switching techniques for NoC »Mesh, fat-tree, honey-comb »Greedy, Deflection, Wormhole, Virtual-Channels n Start with the paper – Design of a Low-Latency Virtual-Channel Router

3. 3 - Sailesh Kumar - 8/8/2015 Why NoC n Billion transistor era has arrived »Several such SoC are in pipeline, Inter-connection is critical n A generic inter-connection architecture ensures »Reduced design time »IP reuse »Predictable backend (versus ad-hoc wiring) n Bus based inter-connects were sufficient until now n 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

4. 4 - Sailesh Kumar - 8/8/2015 Why NoC (cont) n Recently Dally proposed an idea »“Route packets not wires” as in data networks –Point to point communication »Point to point links are faster n 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 n Challenges »Topologies, Routing protocol »Network and router design with small footprint and low latency

5. 5 - Sailesh Kumar - 8/8/2015 Some more motivations n The need to put repeaters into long wires allows us to add the switching needed to implement a network at little additional cost n Makes efficient use of critical global wiring resources by sharing them across different senders and receivers n Simplifies overall design n Design a single router and do copy-paste in both dimension

6. 6 - Sailesh Kumar - 8/8/2015 A typical NoC node n Layered Design of reconfigurable micronetworks. n Exploits methods and tools used for general network. n Micronetworks based on the ISO/OSI model. n NoC architecture consists of Physical, Data link, and Network layers.

7. 7 - Sailesh Kumar - 8/8/2015 A typical NoC n Layered Design of reconfigurable micronetworks. n Exploits methods and tools used for general network. n Micronetworks based on the ISO/OSI model. n NoC architecture consists of Physical, Data link, and Network layers. Implemented in cores, enables end-to- end reliable transport

8. 8 - Sailesh Kumar - 8/8/2015 A typical NoC n Layered Design of reconfigurable micronetworks. n Exploits methods and tools used for general network. n Micronetworks based on the ISO/OSI model. n NoC architecture consists of Physical, Data link, and Network layers. Implemented in cores, enables end-to- end reliable transport Multi hop route setup, packet addressing, etc

9. 9 - Sailesh Kumar - 8/8/2015 A typical NoC n Layered Design of reconfigurable micronetworks. n Exploits methods and tools used for general network. n Micronetworks based on the ISO/OSI model. n NoC architecture consists of Physical, Data link, and Network layers. Implemented in cores, enables end-to- end reliable transport Multi hop route setup, packet addressing, etc Contention issues, reliability issues, grouping of physical layer bits, e.g. “flits”

10. 10 - Sailesh Kumar - 8/8/2015 A NoC topology n Cores Communicates With Each Other Using NoC n NoC Consists of Routers (R) and Network Interfaces (NI) n A NI linked to Router by Non-Pipelined Wires n One or More Cores Connected to a NI

11. 11 - Sailesh Kumar - 8/8/2015 Another NoC topologies Multi hop route setup, packet addressing, etc Fat tree Mesh

12. 12 - Sailesh Kumar - 8/8/2015 Routing protocols n We will only consider mesh topology n Objective is to find a path from a source to a destination n Greedy Algorithms (deterministic) »Choose shortest path (e.g. X-Y) n Adaptive routing »If congestion, choose alternative path »Deflection routing »Is adaptive better than greedy => NOT REALLY (when only local information is used) n Adaptive routing can also result in livelock

13. 13 - Sailesh Kumar - 8/8/2015 Switching techniques n 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. n Store-and-forward policy (Packet Switching): each switch waits for the full packet to arrive in switch before sending to the next switch n 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).

14. 14 - Sailesh Kumar - 8/8/2015 Wormhole Routing – Good fit for NoC n Wormhole routing is good for NoC »Low latency »Less buffering requirements n Suffers from deadlock [example from Li and McKinley, IEEE Computer v26n2, 1993]

15. 15 - Sailesh Kumar - 8/8/2015 Adding Virtual Channels n With virtual channels, deadlock can be avoided n Move message and reply on different channels => Will never have loop on a single channel

16. 16 - Sailesh Kumar - 8/8/2015 Designing Virtual Channel Routers n Design Constraints in NoC »Minimize Latency »Minimize Buffering »Minimal footprint n Can exploit far greater number of pins and wires n May use fat data and flow control wires n Objective: Design routers with minimal latency »This will also result in smaller buffers n This paper presents design of a low latency router »Cycle time of 12 FO4 »Single cycle routing/switching

17. 17 - Sailesh Kumar - 8/8/2015 Designing Virtual Channel Routers n Design Constraints in NoC »Minimize Latency »Minimize Buffering »Minimal footprint n Can exploit far greater number of pins and wires n May use fat data and flow control wires n Objective: Design routers with minimal latency »This will also result in smaller buffers n This paper presents design of a low latency router »Cycle time of 12 FO4 »Single cycle routing/switching

18. 18 - Sailesh Kumar - 8/8/2015 A Virtual Channel Router

19. 19 - Sailesh Kumar - 8/8/2015 Designing Virtual Channel Routers Arriving flits are placed into the buffers of corresponding VC Every VC of every input port has buffers to hold arriving flits

20. 20 - Sailesh Kumar - 8/8/2015 Designing Virtual Channel Routers Arriving flits are placed into the buffers of corresponding VC 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

21. 21 - Sailesh Kumar - 8/8/2015 Designing Virtual Channel Routers Arriving flits are placed into the buffers of corresponding VC 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 Matches successful input ports (allocated VC) to output ports Flits at input VCs getting grants are passed to output VCs

22. 22 - Sailesh Kumar - 8/8/2015 Routing Logic n Three possibilities »Return a single VC »Return set of VCs on a single port »Return any VCs n Look ahead routing »Routing performed at the previous router »Good for X-Y deterministic (non adaptive) routing »A SGI routing chip first implemented it

23. 23 - Sailesh Kumar - 8/8/2015 VC Allocation n Complexity of VC allocation depends on routing range n Routing returns single VC »Needs PxV input arbiter for every outgoing VC n Routing returns multiple VC at single port »Additional V:1 arbiter at every input VC to reduce potential outgoing VC to 1 n Routing returns any set of VCs »Needs two cascaded PxV input arbiters n We consider multiple VC at single port case

24. 24 - Sailesh Kumar - 8/8/2015 VC Allocation Logic »At every outgoing VC following logic is needed

25. 25 - Sailesh Kumar - 8/8/2015 Switch Allocation n Individual flits at input VCs arbitrate for access to the crossbar port n Arbitration can be performed in two stages n First stage »A VC among V possible VCs at every input port is selected »V:1 arbiter at every input port n Second stage »Winning VC at every input port is matched to the output port »P:1 arbiter at every output port n This scheme doesn’t guarantee a maximal/maximum/good matching n But simple to implement

26. 26 - Sailesh Kumar - 8/8/2015 Switch Allocation

27. 27 - Sailesh Kumar - 8/8/2015 Issues n VC allocation and Switch allocation are serialized n A flit will either take 2 clocks to get through n Else clock speed will be low n Solution: Speculative switch allocation

28. 28 - Sailesh Kumar - 8/8/2015 Speculative Switch Allocation n 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 n 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 n Mostly one cycle delay under light load n Mostly one cycle delay under heavy load Speculative will work Non-speculative will work

29. 29 - Sailesh Kumar - 8/8/2015 n Further Enhancement n Is it possible to have zero cycle VC/switch allocation YES, Most of the time, that’s what this paper is about!

30. 30 - Sailesh Kumar - 8/8/2015 Idea 1: Free Virtual Channel Queue n Keep queue of free VC at every outgoing port »Also bit mask with one set bit n Thus First stage of VC allocation where an output VC is selected will be removed

31. 31 - Sailesh Kumar - 8/8/2015 Idea 1: Free Virtual Channel Queue n Keep queue of free VC at every outgoing port »Also bit mask with one set bit n Thus First stage of VC allocation where an output VC is selected will be removed

32. 32 - Sailesh Kumar - 8/8/2015 Idea 2: Pre-computing arbitration decisions n If somehow, you know the arbitration results before flits actually arrive and fight for the VC and switch n I mean, every arbitration decision »VC allocation »Switch allocation »Etc n Then the router can be made to run in zero cycle »The arriving flit route/switch in the same clock they arrive n Also, clock speed may be pretty good »Data path and control path are no more in series n That’s what the idea 2 is.

33. 33 - Sailesh Kumar - 8/8/2015 Some preliminaries before going into detail n Tree Arbiters »Implements large arbiters using tree of small arbiters n Matrix Arbiters »Fair and Fast arbiter implementation

34. 34 - Sailesh Kumar - 8/8/2015 VC allocation using a Tree Arbiter

35. 35 - Sailesh Kumar - 8/8/2015 A Matrix Arbiter

36. 36 - Sailesh Kumar - 8/8/2015 Pre-computing arbitration decisions n An alternative arbiter design

37. 37 - Sailesh Kumar - 8/8/2015 Pre-computing arbitration decisions n An alternative arbiter design Generate grant enables one cycle prior and latch them Grants are product of latched enables and the requests

38. 38 - Sailesh Kumar - 8/8/2015 Pre-computing arbitration decisions n An alternative arbiter design n Grants are generated in same clock as request arrives »If at least one request remains Generate grant enables one cycle prior and latch them Grants are product of latched enables and the requests

39. 39 - Sailesh Kumar - 8/8/2015 Pre-computing arbitration decisions n An alternative arbiter design n However, when no request remains, it is difficult to generate grant enables ??? Generate grant enables one cycle prior and latch them Grants are product of latched enables and the requests

40. 40 - Sailesh Kumar - 8/8/2015 Generating grant enables n 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 n 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 n All first stage V:1 arbiters operate under safe environment n However P:1 arbiters doesn’t

41. 41 - Sailesh Kumar - 8/8/2015 Generating grant enables n 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 n Why will it work? n Because in lightly loaded network, multiple requests for same VC/port will not arrive (few aborts) n In heavily loaded network flits will remain buffered and Non-speculative arbitration (higher priority) will happen most of the time »Few aborts again

42. 42 - Sailesh Kumar - 8/8/2015 I will skip the design details now n Since it is confusing and complex n Will jump to critical path analysis

43. 43 - Sailesh Kumar - 8/8/2015 Analysis of critical path Generates VC/switch grants from pre- computed grant enables

44. 44 - Sailesh Kumar - 8/8/2015 Analysis of critical path Generates VC/switch grants from pre- computed grant enables Crossbar traversal is aborted once invalid grants are detected

45. 45 - Sailesh Kumar - 8/8/2015 Analysis of critical path Generates VC/switch grants from pre- computed grant enables Crossbar traversal is aborted once invalid grants are detected In case of an abort, the correct control signals are ensured in the next cycle

46. 46 - Sailesh Kumar - 8/8/2015 Final design n Control path critical delay is 12 FO4 »Until now, the best design had 20 FO4 delays n They have sampled a NoC based ASIC last week using this idea n Runs at several GHz speeds n 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)

47. 47 - Sailesh Kumar - 8/8/2015 Simulation results

48. 48 - Sailesh Kumar - 8/8/2015 If (doubts) Then Ask; Else Thank you; Goto Discussion; End if;