Towards Energy-Propor1onal Op1cal Interconnects

Transcription

1 Towards Energy-Proporonal Opcal Interconnects Nikos Hardavellas, Northwestern University Yigit Demir, Computa8onal Lithography, Intel OPTICS Workshop March 8 th, 206 par8ally supported by NSF award CCF Photonics Need High Power Lasers Emergence of photonics High bandwidth, low latency, energy efficient Wide range of apps: manycores, mul8-chip, datacenters However, lasers are really power-hungry q Op8cal devices induce op8cal loss (3+ db is typical) q WDM-compa8ble lasers are 5 30% efficient 0 20x higher power than required op8cal output 2

2 Most of the Laser Power is Wasted Demir & Hardavellas [HPCA 5] [NOCS 5] [SPIE 5] [IPC 4] [ISLPED 4] Interconnect may stay idle for long 8mes q Compute-intensive execu8on phases of workloads q 30% server u8liza8on in data centers [Barroso 2007] But laser stays always on! q even during periods of interconnect inac8vity Up to 94% laser energy waste in real-world workloads 3 Proposed Soluon: Laser Power-Gang Turn the lasers off when interconnect is idle Turn the lasers on before sender transmits q This may be tricky needs early warning or predic8ve schemes Overlooked un8l recently q Tradi8onal comb lasers are slow to turn on New enabling technology: Fast on-off switching on-chip lasers q InP, Ge, Turn on/off in.5 2 ns q On-chip à simplify design and lower cost q [HPCA 5] [SPIE 5] [IPC 4] [ISLPED 4] 4

3 ProLaser: Energy-Proporonal Photonic Nets. Power saving mechanism for photonic interconnects q Laser power-ga8ng à Independent power ga8ng for data and control bits à Predicts laser turn-on à Saved power can be used by the cores Result highlights q Laser energy reduc8on: 42 88% (6% on avg.) q Processor energy reduc8on: 35 52% (40% on avg.) q Leads to 50 73% speedup (60% on avg.) q Within 2 6% of the theore8cally maximum savings 5 Architecture Assumed: Tiled Mulcore Dir Router Dir Router LLC slice LLC slice Tile Core, L Tile Core, L Dir Router Dir Router LLC slice LLC slice Tile Core, L Tile Core, L Dir Router Dir Router LLC slice LLC slice Tile Core, L Tile Core, L 6 Dir Router LLC slice Tile Core, L Dir Router LLC slice Tile Core, L Dir Router LLC slice Tile Core, L 6

4 Architecture Assumed: R-SWMR Opcal Crossbar Router 0 (Home) Data Bus D D 0 R R 0 Reservation Channel Router Router N-2 Router N- 7 Segregang Data from Control Bits Most on-chip messages are short. On average: 65% are Control Messages 88 bits 35% are Data Messages 600 bits WDM Laser WDM Laser WDM Laser WDM Laser WDM Laser WDM Laser Turn on the common bits (44) for all types of data Turn on the data-only bits (300) for data messages only 8 [LaC, Demir & Hardavellas, IPC 4] λ λ 2 λ N λ λ 2 λ N λ λ 2 λ N λ λ 2 λ N λ λ 2 λ N λ λ 2 λ N Data-Only Bits Common Bits Data Bus

5 Predicve Laser Turn-On: Communicaon Iniaon cycle Bloom LLC slice 0+ cycles Bloom filter monitors LLC q LLC access: ~4 cycles q LLC tag lookup: ~0 cycles q Bloom filter: cycle Tile Core, L 9 KB coun8ng Bloom filter q <2% false posi8ve Laser Control Co-design with Coherence Protocol [EcoLaser+, Demir & Hardavellas, SPIE 5] Forwarded Request Request Bloom LLC slice Directory Core, L Ack Bloom LLC slice Reply Bloom LLC slice Requestor Core, L 0 Owner Core, L

6 Laser Control Co-design with Coherence Protocol [EcoLaser+, Demir & Hardavellas, SPIE 5] An8cipates laser ac8va8on q Correlates cache coherence requests to replies q Ac8vates laser early à hides laser turn-on delay Which laser / plane to turn on? q Predict cache miss à turn on requestor s control plane q Request to directory à turn on directory s control plane q Directory forwarding à turn on owner s control+data plane q etc (including memory controller) When to turn it on? q Turn-on the laser just.5ns before the payload is ready q Minimum latency for each opera8on Router Microarchitecture (R-SWMR) Reserva8on Channels Data Channels L2 Cache Requests & Replies L L L2 Cache Slice Bloom Filter L L R R 2 R N L L RCH N RCH 2 RCH CH CH 2 CH N Inject Inject C Data Channel O/E O/E VC0 VC VC2 VC0 VC VC2 Data Channel N Switch Allocator & VC Allocator E/O E/O Laser Controller E/O Eject Eject C Lasers Data Channel i Common Channel i Reserva8on Channel i Bloom filters + coherence protocol à predict accesses 2

7 Controlling Off-chip Laser Source Off-chip laser die Network-on-chip [Heck & Bowers, JSTQE 4] WDM WDM WDM WDM WDM WDM λ λ 2 λ N λ λ 2 λ N Laser Switch λ λ N Op8cal Fiber λ λ N λ & λ 3 λ λ 2 λ N λ λ 2 λ N SOI Waveguides λ λ 2 λ N λ λ 2 λ N Data-Only Bits Common Bits Data Bus Messages CDF Interconnect OSen Stays Idle.0 Bodytrack 0.8 Em3d Ocean 0.6 Appbt Tomcatv 0.4 Barnes 0.2 Moldyn FMM 0.0 Average Message Inter-Arrival Time (cycles) 70% messages on avg. 9 cycles apart (.8ns), 40% are 4ns 4

8 Norm. Energy / Instr ProLaser Energy Savings Core_Leakage Core_Dynamic Memory_Leakage Memory_Dynamic Ring_Heating Modulation Laser Electrical Network F N e S E P I N e S E P I... F N e S E P I N e S E P I F N e S E P I N e S E P I On-chip Off-chip On-chip Off-chip On-chip Off-chip... Fmm Bodytrack Average ProLaser saves 49 88% of laser power 35 52% lower energy / instruc8on (40% avg.) 5 ProLaser: Performance Impact on Real Workloads Speedup Flat-Butterfly No-Ctrl Average Simple EcoLaser ProLaser Perfect No-Ctrl-OffChip Power_Eq-OffChip Simple-OffChip EcoLaser-OffChip ProLaser-OffChip Perfect-OffChip 60% speedup over No-Ctrl; 40% over flazened buzefly 6

9 Teaser Slide: Laser Gang in the Datacenter Laser Energy/Flit No-Ctrl Naive SLAC SLAC w/off EDU EDU2 AVERAGE SLaC w/off op8miza8on saves 79% of the laser energy Similar results with Facebook and Microso{ traces 7 Conclusion Problem: lasers are really power hungry, mostly wasted power Our solu8on: laser power-ga8ng (ProLaser, SLaC, EcoLaser (+), LaC) Significant energy reduc8on q Laser: 42 88% (6% on avg.), Processor: 35 52% (40% on avg.) q Within 2 6% of the theore8cally maximum savings q Power reduc8on leads to speedups: 50 73% (60% on avg.) Applicable to a wide range of scales (on chip, mul8chip, datacenter) Thank you! Quesons? 8

10 Backup Slides 9 DTM for Temperature Limi8ng Simulaon Tool Chain Cycle Accurate Full System Simula8on Flexus 4.0 Booksim 2.0 DRAMSim Run8me Stats Power Calcula8ons McPat 0.8 DSENT + Analycal Model Opera8ng Temperature Leakage and Dynamic Power HotSpot

11 Experimental Methodology CMP Size 64 cores, 580 mm 2 Core L Cache L2 Cache Memory Controller Main Memory Network ULTRASPARC III ISA, up to 5Ghz, OoO, 4-wide dispatch/retirement, 96-entry ROB Split I/D, 64KB 2-way, 2-cycle load-to-use, 2 ports, 64-byte blocks, 32 MSHRs, 6-entry victim cache 52 KB per core, 6 way, 64-byte blocks, 4 cyclehit, 32 MSHRs, 6-entry victim cache One per 4 cores, channel per Memory Controller, Round-robin page interleaving Optically connected memory [3], 0ns access R-SWMR radix-6 crossbar and firefly, 300-bit wide 0GHz, 20 flit deep buffers, 3 cycle router delay 2 Nanophotonic Parameters 22

12 Fmm: Input 28K Moldyn: 5, 20, 3.2 M Barnes: Input 64K Tomcatv: 4096, 0 Appbt: in.24x24x24x8bit Ocean: 026, 9600 Em3d: 400K, 2, 5, 5 Bodytrack Workloads 23 Latency (cycles) Interconnect Performance No-Control Perfect EcoLaser [Demir & Hardavellas, ISLPED 4] EcoLaser-OffChip Simple Simple-OffChip ProLaser ProLaser-OffChip Injection Rate ProLaser almost perfect satura8on; EcoLaser saturates early 24

13 Norm. Energy / Flit Interconnect Energy No-Control-OffChip EcoLaser-OffChip LaC-OffChip ProLaser-OffChip Injection Rate ProLaser saves 49 88% of laser power ProLaser is ~2x bezer than EcoLaser; 2 6% of Perfect 25 No-Control EcoLaser LaC ProLaser Sensivity to Laser Turn-On Delay EcoLaser - LES ProLaser - LES ProLaser-w/o Bloom Filters No-Ctrl ProLaser-w/o Bloom Filters - LES EcoLaser ProLaser Laser Energy Savings (LES) 00% 75% 50% 25% 0% Laser Turn-on Latency (ns) ProLaser tolerates 2.3x higher laser turn-on delay than EcoLaser Average Latency (cycles)

14 Speedup (On-chip Lasers) Flat-Butterfly No-Ctrl Power Eq Simple EcoLaser ProLaser Perfect Fmm Moldyn Barnes Tomcatv Appbt Ocean Em3d Bodytrack Average Speedup Speedup (Off-chip Lasers) Flat-Butterfly No-Ctrl-OffChip Power_Eq-OffChip Simple-OffChip EcoLaser-OffChip ProLaser-OffChip Perfect-OffChip 2 Speedup Fmm Moldyn Barnes Tomcatv Appbt Ocean Em3d Bodytrack Average

15 Router microarchitecture for SWMR EcoLaser [Ecolaser, Demir & Hardavellas, ISLPED 4] Message in injec8on buffers à Laser Turn On 29 Router Microarchitecture for MWSR EcoLaser [Ecolaser, Demir & Hardavellas, ISLPED 4] Adap8ve Laser Turn-on + Specialized Token Stream 30

16 MWSR Opcal Bus Data Bus D D 0 T T 0 Token Stream Router 0 Router Router N-2 Router N- (Home) 3