1. Hybrid-Scheduling: A Compile-Time Approach for Energy–Efficient Superscalar Processors Madhavi Valluri and Lizy John Laboratory for Computer Architecture University of Texas at Austin www.ece.utexas.edu/projects/ece/lca Loop: { (1) add R1, R2, R5; (2) sub R5, R2, R1; (3) st R1, 0(R15); (4) ld R9, 8(R24); (5) add R4, R9, R10; } Compiler Schedule (case 1) Cycle 1 Instr (1) Cycle 2 Instr (2) Cycle 3 Instr (3) SIX CYCLES Cycle 4 Instr (4) PER ITERATION Cycle 5 nop Cycle 6 Instr (5) Dynamic Schedule Cycle 1 Instr (1) Instr (4) Cycle 2 Instr (2) nop Cycle 3 Instr (3) Instr (5) THREE CYCLES PER ITERATION Dynamic schedule is 2X better than compiler schedule Assumed operation latencies ALU – 1 cycle LD/ST – 2 cycles Compiler Schedule (case 2) Cycle 1 Instr (1) Instr (4) Cycle 2 Instr (2) nop Cycle 3 Instr (3) Instr (5) THREE CYCLES PER ITERATION CASE 1 : Assume R15 & r24 depend on input data; Potential address aliasing; cannot be resolved statically CASE 2: R15 and R24 are starting addresses of two arrays a & b; no aliasing problem exists Motivating Example Where does power go? Where does power go? Out-of-order issue logic* used to identify parallel instructions to issue each cycle consumes ~25-50% of processor power OoO Issue Logic Main Components: Issue Queue & Reorder Buffer - Accessed multiple times each cycle - Highly Associative - Multiple locations accessed - Large number of ports - Complex Select/Wakeup logic Alpha 21464 The Hybrid-Scheduling Approach The Hybrid-Scheduling Approach Preliminary Results (ISLPED 2003) Programs divided into low power static regions and high power dynamic region Code in S-Regions scheduled by compiler alone into packets of independent instructions Instruction packets issued to functional units without any dynamic dependence checks 2 modes of execution – Superscalar mode for dynamic regions + Static mode for S-Regions Large savings in static mode since OoO logic bypassed Benchmarks & S-Region Selection Media & SpecFP benchmarks Potential S-Regions - Loops without function calls Final S-Regions selected by profiling 30-99% time spent in S-Regions Comparing with Dynamic Resource Adaptation Schemes Resource requirements of program varies with program phase Dynamic resource adaptation schemes typically lower processor resources when program IPC is low and increase resources when IPC is high Hybrid-Scheduling scheme eliminates power-hungry resource use for high IPC (or high ILP) regions Future Research Directions Future Research Directions Experimental Results 30% average energy savings 3.6% average performance drop NO difference between dynamic schedule and compiler schedule! Decode Rename O-o-O Issue Reorder Buffer alu Low Power Reorder Buffer Special instruction to switch mode Program Dynamic Region Static Region a.k.a: Dynamic Issue Logic Fetch FPU IALU Caches dcache Rest OoO issue Exec Mem Fetch OoO issue OoO issue Rest IALU FPU MMU Dynamic scheduling hardware justified in irregular regions of the program In regular regions of programs (regions with well-structured control-flow, regular memory access patterns etc), hardware scheduling is redundant SOLUTION: Hybrid-Scheduling Approach The real challenge: Adapting Hybrid-scheduling for irregular, SPECint-like applications Minimal time spent in loops without function calls (previously explored type of S-Region) Ongoing research – Investigating alternate S-Regions – Hyperblocks, Superblocks and Traces