Real-Time Communication Analysis for NoCs with Wormhole Switching NoCs 2 0 0 8 Zheng Shi and Alan Burns Real-time system group Department of computer science The University of York
Outline 1 QoS in NoC 2 Priority based w orm hole sw itching 3 W orst case netw ork latency analysis 4 Conclusion LOGO Your site here
Netw orks on Chip ( NoC) � On-chip Communication: � Point-to-Point � Bus � NoC: packet-switched, shared, optimized for communications � Resource efficiency � High scalability � IP reusability � High performance LOGO Your site here
NoC needs QoS � Differentiated Service Requirement � Best Effort � Guaranteed Service � Performance parameters: latency, bandwidth, bounded jitter and loss probability, in- order data, etc. � Real-Time Service: � The correctness relies on not only the communication result but also the completion time bound (deadline). � For hard real-time service, it is necessary that all the packets must be delivered before their deadlines even under worst case scenario. LOGO Your site here
Several Solutions � Contradiction : The network gives more efficiency and flexibility but introduces the unpredictable delay due to the contention. Real-time service, requires the timing to be predictable even under the worst case situation � Contention avoidable � Circuit Switching : aSoC � TDM : AEtheral, Nostrum � Contention acceptable � Priority based Wormhole Switching LOGO Your site here
W orm hole Sw itching � Advantages (with Virtual Channels) � Small Buffer Size � High Throughput � Low Average Latency LOGO Your site here
Priority Router Structure � There are sufficient VCs at each router � Each VC is assigned distinct global priority � Each flow also has distinct priority � Flow only requests the VC with same priority � At any time, only the flit with highest priority can access the output link � Flit-level priority preemption between different VCs LOGO Your site here
System Model � Characterize traffic-flow � A traffic-flow is packet stream which traverses the same route from source to destination and requires the same grade of service. � Attribute � P : Priority � C : Basic network latency � T : Period for periodic flow or minimal interval for sporadic flow � D : Deadline R : Release Jitter J � � Interrelationship τ ∩ τ ≠ φ � Direct competing: ( ) ( ) Path Path i j = ∀ τ τ ∩ τ ≠ φ > direct interference set: D { | ( ) ( ) , } S Path Path P P i j i j j i � Indirect competing: τ ∩ τ ≠ φ τ ∩ τ ≠ φ τ ∩ τ = φ ( ) ( ) , ( ) ( ) , ( ) ( ) Path Path Path Path Path Path i j j k i k indirect interference set = ∀ τ τ ∩ τ ≠ φ τ ∩ τ ≠ φ τ ∩ τ = φ > > I { | ( ) ( ) , ( ) ( ) , ( ) ( ) , } S Path Path Path Path Path Path P P P i k i j j k i k k j i LOGO Your site here
W orm hole Sw itching- A Case τ τ 1 3 τ 2 = φ = φ D I , Priority ordering: S S 1 1 > > P P P = τ = φ D I { }, S S 1 2 3 2 1 2 = τ = τ D I { }, { } S S 3 2 3 1 LOGO Your site here
Characterize Netw ork Latency � Worst case network latency R : � The maximum length of time the packet could take to travel from source to destination R ≤ � The flow is schedulable if D � Basic network latency C : the network latency happens when there no traffic-flow contention exists. ⎡ ⎤ + L L = ⋅ + ⋅ max add ⎢ ⎥ / C f B Hop S size link ⎢ ⎥ f size LOGO Your site here
Model and Assum ption � The physical communication links are treated as shared competition resource � At any time, only one traffic-flow is permitted to access the shared path � The packet moves ahead when gets highest priority along the path � The arrivals of higher priority flows are considered as preemption interference � The allowable service time for a flow is all the time interval at which no higher priority flow competes for the same physical link LOGO Your site here
Netw ork Latency Evaluation( 1 ) � Worst Case Network Latency: = + R C I i i i R : worst case latency i I : maximum interference i the packets is supposed with maximum length and released at maximum rate ⎡ ⎤ + R R J ∑ = i j ⎢ ⎥ I C i j ⎢ ⎥ T ⎢ ⎥ ∀ ∈ D j S j i LOGO Your site here
Netw ork Latency Evaluation( 2 ) � Worst case network latency equation ⎡ ⎤ + R ∑ R J = + ⎢ ⎥ i i R C C i i j ⎢ ⎥ T ⎢ ⎥ ∀ ∈ D j S j i The eqaution is solved using iterative technique ⎡ ⎤ + n R R J ∑ + = + n 1 ⎢ ⎥ i i R C C i i j ⎢ ⎥ T ⎢ ⎥ ∀ ∈ D j S j i = 0 Iterative starts with R C i i + 1 = and terminates when n n R R i i + 1 > n or which denotes the deadline miss for this flow. R D i i LOGO Your site here
Consider I ndirect I nterference ( 1 ) � Minimal interval between subsequent preemption is less than period τ 1 T i τ 2 R i τ 3 � This could happen only when indirect interference is considered. LOGO Your site here
Consider I ndirect I nterference ( 2 ) � Preemption interference upper bound T j T T T j j j R j τ j C R I J J j j j R i τ i ⎡ ⎤ + + − R R J R C ∑ = i j j j ⎢ ⎥ I C i j ⎢ ⎥ T ⎢ ⎥ ∀ ∈ D j S j i � Worst case latency ⎡ ⎤ + + − R R J R C ∑ = + i i j j ⎢ ⎥ R C C i i j ⎢ ⎥ T ⎢ ⎥ ∀ ∈ D j S j i LOGO Your site here
Case Exam ple τ � Trafffic- C P T D For : there is no higher priority flow 1 τ Flow s = C = than , so 2 R 1 1 1 τ 2 1 6 6 1 τ τ � τ For : shares the physical link with 2 3 2 7 7 2 τ = τ = φ D I 2 { }, higher priority flow and S S 1 2 1 2 τ 3 3 13 13 = 0 3 R 3 2 τ ⎡ ⎤ 3 = + = � 1 suffers both direct and indirect 3 2 5 R ⎢ ⎥ 3 2 ⎢ ⎥ 6 = τ = τ D I { }, { } interference with S S 3 2 3 1 ⎡ ⎤ 5 = + = 2 3 2 5 R ⎢ ⎥ τ 2 ⎢ ⎥ 6 The interference jitter of referred to 2 τ − C = − = equals 5 3 2 R 3 2 2 So ⎡ ⎤ + − R R C = + 3 2 2 ⎢ ⎥ R C C 3 3 2 ⎢ ⎥ T 2 3 = 9 which stops at R LOGO Your site here
Tightness of analysis ( 1 ) τ τ 1 2 τ 3 = φ = φ D I , S S Priority ordering: 1 1 > > = φ = φ P P P D I , S S 1 2 3 2 2 = τ τ = φ D I { , }, S S 3 1 2 3 LOGO Your site here
Tightness of analysis ( 2 ) � Parallel Interference τ 1 τ 2 τ 3 � When parallel interference exists, the real worst case network latency is no more than the analysis result. � When parallel interference exists, finding worst case network latency is NP-hard (the proof details refers the paper). � Our analysis is safe but pessimistic. LOGO Your site here
Conclusion � Real time communication service can be supported by priority based wormhole switching technique. � The schedulable test is derived by worst case network latency analysis. � Both direct and indirect interferences are taken into account. � When parallel interference exists, finding worst case network latency is NP-hard, but our analysis still form an upper bound. LOGO Your site here
and Question Thank you…
Recommend
More recommend
