Fine-Grained Bandwidth Adaptivity in Networks-on-Chip Using - PowerPoint PPT Presentation

Fine-Grained Bandwidth Adaptivity in Networks-on-Chip Using Bidirectional Channels Robert Hesse , Jeff Nicholls, Natalie Enright Jerger University of Toronto May 10, 2012 Friday, 11 May, 12

Motivation May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation • NoCs are crucial for scaling CMPs May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation • NoCs are crucial for scaling CMPs • Problem: May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation • NoCs are crucial for scaling CMPs BW • Problem: – NoC bandwidth resources are static Time May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation • NoCs are crucial for scaling CMPs BW • Problem: – NoC bandwidth resources are static – Bandwidth requirements are highly Time dynamic May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation • NoCs are crucial for scaling CMPs BW • Problem: – NoC bandwidth resources are static – Bandwidth requirements are highly Time dynamic • Current solution: May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation • NoCs are crucial for scaling CMPs BW • Problem: – NoC bandwidth resources are static – Bandwidth requirements are highly Time dynamic • Current solution: – Over-provisioned link BW May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation • NoCs are crucial for scaling CMPs BW • Problem: – NoC bandwidth resources are static – Bandwidth requirements are highly Time dynamic • Current solution: Average channel utilization: < 5% – Over-provisioned link BW May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation • NoCs are crucial for scaling CMPs BW • Problem: – NoC bandwidth resources are static – Bandwidth requirements are highly Time dynamic • Current solution: Average channel utilization: < 5% – Over-provisioned link BW • Our solution: May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation • NoCs are crucial for scaling CMPs BW • Problem: – NoC bandwidth resources are static – Bandwidth requirements are highly Time dynamic • Current solution: Average channel utilization: < 5% – Over-provisioned link BW • Our solution: – Adapt link BW to demands May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation • NoCs are crucial for scaling CMPs BW • Problem: – NoC bandwidth resources are static – Bandwidth requirements are highly Time dynamic • Current solution: Average channel utilization: < 5% – Over-provisioned link BW • Our solution: Save up to 75% of BW resources – Adapt link BW to demands May 10, 2012 2 University of Toronto Friday, 11 May, 12

Motivation - Static NoC May 10, 2012 University of Toronto 3 Friday, 11 May, 12

Motivation - Static NoC IP IP IP IP 0 1 2 3 IP IP IP IP 4 5 6 7 IP IP IP IP 8 9 10 11 IP IP IP IP 12 13 14 15 • Static Topology May 10, 2012 University of Toronto 3 Friday, 11 May, 12

Motivation - Static NoC IP IP IP IP 0 1 2 3 IP IP IP IP 4 5 6 7 IP IP IP IP 8 9 10 11 IP IP IP IP 12 13 14 15 • Static Topology • Static Bandwidth May 10, 2012 University of Toronto 3 Friday, 11 May, 12

Motivation - Static NoC Uniform Random IP IP IP IP 0 1 2 3 IP IP IP IP 4 5 6 7 IP IP IP IP 8 9 10 11 IP IP IP IP 12 13 14 15 • Static Topology • Static Bandwidth • Static workloads for evaluation May 10, 2012 University of Toronto 3 Friday, 11 May, 12

Motivation - Static NoC Uniform Random 15 14 IP IP IP IP IP IP IP IP 13 12 0 0 1 1 2 2 3 3 11 10 IP IP IP IP IP IP IP IP 9 4 4 5 5 6 6 7 7 8 7 IP IP IP IP IP IP IP IP 6 8 8 9 9 10 10 11 11 5 4 3 IP IP IP IP IP IP IP IP 2 1 12 12 13 13 14 14 15 15 0 • Static Topology • Static Bandwidth • Static workloads for evaluation May 10, 2012 University of Toronto 3 Friday, 11 May, 12

Motivation - Static NoC Transpose 15 14 IP IP IP IP IP IP IP IP 13 12 0 0 1 1 2 2 3 3 11 10 IP IP IP IP IP IP IP IP 9 4 4 5 5 6 6 7 7 8 7 IP IP IP IP IP IP IP IP 6 8 8 9 9 10 10 11 11 5 4 IP IP IP IP IP IP IP IP 3 2 12 12 13 13 14 14 15 15 1 0 • Static Topology • Static Bandwidth • Static workloads for evaluation May 10, 2012 University of Toronto 3 Friday, 11 May, 12

Motivation - Static NoC Transpose 15 14 IP IP IP IP IP IP IP IP 13 12 0 0 1 1 2 2 3 3 11 10 IP IP IP IP IP IP IP IP 9 4 4 5 5 6 6 7 7 8 7 IP IP IP IP IP IP IP IP 6 8 8 9 9 10 10 11 11 5 4 IP IP IP IP IP IP IP IP 3 2 12 12 13 13 14 14 15 15 1 0 • Specified at design time • Static Topology for worst case scenario • Static Bandwidth • Static workloads for evaluation May 10, 2012 University of Toronto 3 Friday, 11 May, 12

Motivation - Static NoC Transpose 15 14 IP IP IP IP IP IP IP IP 13 12 0 0 1 1 2 2 3 3 11 10 IP IP IP IP IP IP IP IP 9 4 4 5 5 6 6 7 7 8 7 IP IP IP IP IP IP IP IP 6 8 8 9 9 10 10 11 11 5 4 IP IP IP IP IP IP IP IP 3 2 12 12 13 13 14 14 15 15 1 0 • Specified at design time • Static Topology for worst case scenario • Static Bandwidth • Static NoCs can handle • Static workloads for temporally- and spatially- evaluation stable traffic well May 10, 2012 University of Toronto 3 Friday, 11 May, 12

Motivation - Real NoC Tra ffj c Blackscholes 15 IP IP IP IP 14 13 0 1 2 3 12 11 IP IP IP IP 10 9 4 5 6 7 8 7 IP IP IP IP 6 8 9 10 11 5 4 IP IP IP IP 3 2 12 13 14 15 1 0 May 10, 2012 University of Toronto 4 Friday, 11 May, 12

Motivation - Real NoC Tra ffj c Blackscholes 15 IP IP IP IP 14 13 0 1 2 3 12 11 IP IP IP IP 10 9 4 5 6 7 8 7 IP IP IP IP 6 8 9 10 11 5 4 IP IP IP IP 3 2 12 13 14 15 1 0 • Highly dynamic workloads May 10, 2012 University of Toronto 4 Friday, 11 May, 12

Motivation - Real NoC Tra ffj c Blackscholes 15 IP IP IP IP 14 13 0 1 2 3 12 11 IP IP IP IP 10 9 4 5 6 7 8 7 IP IP IP IP 6 8 9 10 11 5 4 IP IP IP IP 3 2 12 13 14 15 1 0 • Highly dynamic workloads • Large temporal and spatial BW variance May 10, 2012 University of Toronto 4 Friday, 11 May, 12

Motivation - Real NoC Tra ffj c Streamcluster 15 IP IP IP IP 14 13 0 1 2 3 12 11 IP IP IP IP 10 9 4 5 6 7 8 7 IP IP IP IP 6 8 9 10 11 5 4 IP IP IP IP 3 2 12 13 14 15 1 0 • Highly dynamic workloads • Large temporal and spatial BW variance May 10, 2012 University of Toronto 4 Friday, 11 May, 12

Motivation - Real NoC Tra ffj c Streamcluster 15 IP IP IP IP 14 13 0 1 2 3 12 11 IP IP IP IP 10 9 4 5 6 7 8 7 IP IP IP IP 6 8 9 10 11 5 4 IP IP IP IP 3 2 12 13 14 15 1 0 • Highly dynamic • Significant area and power overhead with traditional workloads NoC implementation • Large temporal and spatial BW variance May 10, 2012 University of Toronto 4 Friday, 11 May, 12

Motivation - Real NoC Tra ffj c Streamcluster 15 IP IP IP IP 14 13 0 1 2 3 12 11 IP IP IP IP 10 9 4 5 6 7 8 7 IP IP IP IP 6 8 9 10 11 5 4 IP IP IP IP 3 2 12 13 14 15 1 0 • Highly dynamic • Significant area and power overhead with traditional workloads NoC implementation • Large temporal and • Channels are underutilized spatial BW variance most of the time May 10, 2012 University of Toronto 4 Friday, 11 May, 12

Channel Utilization May 10, 2012 5 University of Toronto Friday, 11 May, 12

Channel Utilization 45" 40" Max."U?liza?on" Channel'U)liza)on'(%)' 35" Avg."U?liza?on" 30" 25" 20" 15" 10" 5" 0" " " " " " " " " " " " s k l m t e s s . e r a g e e e n p c t c e v i l a r a t o i a o s n r s V A m r e r e ? u h t n t c F p y i y l c a a n c d a s a C F m a k o R w d c B a S a i u e l B l r F t S May 10, 2012 5 University of Toronto Friday, 11 May, 12

Channel Utilization 45" 40" Max."U?liza?on" Channel'U)liza)on'(%)' 35" Avg."U?liza?on" 30" 25" 20" 15" 3.42% 10" 5" 0" " " " " " " " " " " " s k l m t e s s . e r a g e e e n p c t c e v i l a r a t o i a o s n r s V A m r e r e ? u h t n t c F p y i y l c a a n c d a s a C F m a k o R w d c B a S a i u e l B l r F t S May 10, 2012 5 University of Toronto Friday, 11 May, 12