Chapter 7 Slicing and Dicing

Transcription

1 1/ 22 Chapter 7 Slicing and Dicing Lasse Harju Tampere University of Technology lasse.harju@tut.fi

2 2/ 22 Concentrators and Distributors Concentrators Used for combining traffic from several network nodes Often used with terminals that have bursty traffic characteristics T1 T2 T3 T4 b T b T b T b T b N T0 T1 T2 T3 T4 T5 T6 T

3 3/ 22 Concentrators and Distributors Distributors Distribute traffic from a single node to several network channels Used when the traffic from a node is too large to be handled by a single network channel Distributors increase the serialization latency and reduces load balance b T b T T b N b T To network ports b T

4 4/ 22 Slicing and Dicing Slicing Distribution of a network node to multiple modules (chips or boards) Three approaches 1. Bit slicing 2. Dimension slicing 3. Channel slicing

5 5/ 22 Slicing and Dicing Bit slicing The bits of each channel are divided evenly across the modules Requires extra control Fault recovery and error detection become more complicated wo[7:0] ei[7:0] si[7:0] si[7:4] no[7:0] no[7:4] Network node [7:0] wi[7:0] eo[7:0] wo[7:4] ei[7:4] Network node [7:4] wi[7:4] eo[7:4] so[7:0] ni[7:0] wo[3:0] ei[3:0] Network node [3:0] wi[3:0] eo[3:0] so[7:4] ni[7:4] so[3:0] ni[3:0] si[3:0] no[3:0]

6 6/ 22 Slicing and Dicing Dimension slicing Channels of the node are divided evenly across the modules No control of error detection issues wo[7:0] ei[7:0] si[7:0] si[7:0] no[7:0] no[7:0] Network node [7:0] wi[7:0] eo[7:0] Network node north/south so[7:0] ni[7:0] wo[7:0] ei[7:0] Network node east/west wi[7:0] eo[7:0] so[7:0] ni[7:0]

7 7/ 22 Slicing and Dicing Channel slicing Each channel is split into independend channels and divided across the modules No communication between the slices Distributors are needed to split the incomming channels wo[7:0] ei[7:0] si[7:0] si[7:4] no[7:0] no[7:4] Network node [7:0] wi[7:0] eo[7:0] wo[7:4] ei[7:4] Network node [7:4] wi[7:4] eo[7:4] so[7:0] ni[7:0] wo[3:0] ei[3:0] Network node [3:0] wi[3:0] eo[3:0] so[7:4] ni[7:4] so[3:0] ni[3:0] si[3:0] no[3:0]

8 8/ 22 Chapter 8 Routing Basics Lasse Harju Tampere University of Technology lasse.harju@tut.fi

9 9/ 22 Routing Basics Routing involves selecting a path from a source node to a destination node in a network Routing determines to a large extent how much of the network potential is realized Characteristics of a good routing algorithm Uses the whole potential of the network topology Balances load across the network channels Keeps path length as short as possible (number of hops, latency) Manages a faulty network

10 10/ 22 Routing Examples Examples of routing algorithms: Greedy always send the packet to the direction resulting in the shortest path Uniform random randomly pick the direction of the packet Weighted random randomly pick the direction of the packet but with different probabilities for each direction Adaptive send the packet in the direction for which the local channels has the lowest load

11 11/ 22 Routing Taxonomy Routing algorithms can be classified in terms of how they select the route from source to destination Deterministic - always chooses the same path between a particular source and a particular destination Oblivious - chooses the route without considering the present state of the network (includes deterministic routing as a subset) Adaptive - adapts to the present state of the network A routing algorithm can be represented as a routing relation and a selection function The routing relation returns a set of potential paths between the source and the destination, and the selection function selects the most appropriate one The algorithms can be executed in two ways: all-at-once or incrementally

12 12/ 22 Deterministic Routing Deterministic routing algorithms always select the same route between a particular source destination pair The simplest type of routing algorithms Easy to implement Predictable performance The lack of path diversity can create severe load imbalances

13 13/ 22 Destination-Tag Routing in Butterfly Networks Each digit of the address is used in turn to select the output port at each step of the route ( 1 lower route, 0 upper route) Routing from 3 to 5 in a 2-ary 3-fly network

14 14/ 22 Dimension-Order Routing in Cube Networks Each digit of the destination address is used to compute a preferred direction Routing from 03 to 22 in a 6-ary 2-cube network

15 15/ 22 Chapter 9 Oblivious Routing Lasse Harju Tampere University of Technology lasse.harju@tut.fi

16 16/ 22 Oblivious Routing Oblivious routing does not consider the state of the network Simple algorithm Easy to implement Easy to analyze A tradeoff between locality and imbalance Valian t randomized routing Good load balancing at the expense of locality Minimal oblivious routing Preserves locality

17 17/ 22 Valiant s Randomized Routing The routing is divided into two phases The packet is first sent to a randomly selected intermediate node From the intermediate node the packet is routed to the destination Benefits Excellent load balancing Good performance on worst-case traffic patterns on cube networks Drawbacks Increased latency baused by the two phase operation Poor performance on local traffic

18 18/ 22 Valiant s Randomized Routing on a Torus Routing from 00 to 12 in a 6-ary 2-cube network The intermediate node is selected randomly

19 19/ 22 Minimal Oblivious Routing Minimal oblivious routing attempts to achieve the load balance of randomized routing without giving up the locality The routing is divided into two phases The packet is first sent to an intermediate node The itermediate node is selected randomly from a minimal quadrant From the intermediate node the packet is routed to the destination Benefits Better load balancing than deterministic routing Better performance on local traffic Drawbacks Worse performance on worst-case traffic patterns than randomized routing

20 20/ 22 Minimal Oblivious Routing on a Torus Routing from 00 to 12 in a 6-ary 2-cube network The minimal quadrant is the smallest subnetwork that contains the source and the destination as corner nodes

21 21/ 22 Load-Balanced Oblivious Routing A compromise between Valiant s algorithm and minimal oblivious routing Same routing algorithm as minimal oblivious routing but the location of the quadrant is selected randomly Benefits Outperforms Valiant s algorithm on local traffic Better performance on wors-case traffic patterns

22 22/ 22 Summary Slicing can be used to distribute network nodes to multiple modules Bit slicing, dimension slicing, channel slicing Routing determines how much of the networks potential is utilized A good routing algorithm uses the whole potential of the network topology, balances the traffic, and keeps routes as short as possible Three basic types of routing algorithms: deterministic, oblivious, and adaptive Deterministic routing is the simplest Valiant s routing algorithm performs well on worst-case traffic patterns Minimal oblivious routing preserves locality Load-balanced oblivious routing is the compromise between Valiant s alorithms and minimal oblivious routing