GRAPE — gem5 aRea And Power Estimation

Purpose

gem5 already estimates power for its objects via its own built-in power model, but it isn't wired up by default, and gem5's model alone doesn't cover everything (e.g. Garnet's on-chip network isn't a "cache", so DSENT still does better there, and cache circuit-level numbers from Cacti are more detailed than a hand-written power expression). The goal of this project is to make gem5-based architecture power/area modeling for a RISC-V MinorCPU + Ruby (+ optionally Garnet) system easy to get real numbers out of -- both by wiring gem5's own power model into a running simulation, and by driving external tools (Cacti, DSENT) against a completed run -- working toward eventually calling Cacti directly from gem5's C++ code so its numbers show up next to gem5's own stats without a separate script step.

Current state

• ext/cacti — HewlettPackard/cacti, readonly HTTPS submodule
• ext/gem5 — gem5/gem5 (stable branch), readonly HTTPS submodule
• activate_environment.sh — sourceable setup script that builds and caches Cacti, gem5, and DSENT (see below)
• scripts/run_minor_ruby_power.py — SE-mode run script (fork of gem5's deprecated configs/deprecated/example/se.py) that wires gem5's native MathExprPowerModel onto the CPU and Ruby cache controllers, using real CMOS-shaped power equations (P_dyn = C·V²·f, P_static = V·I_leak0·2^((T-25)/10)), so stats.txt reports actual per-power-state Watt figures instead of only power-state residency ticks. Works with any classic --cpu-type/--ruby combination se.py supports, including RiscvMinorCPU + Ruby + --network garnet.
• scripts/garnet_power_from_m5out.py — for runs that used Garnet as the Ruby network, estimates router/link power and area via DSENT (gem5's own NoC power model). A Python-3-clean rewrite of gem5's util/on-chip-network-power-area.py, which targets Python 2 and doesn't run as-is; see scripts/dsent-py3-patch/README.md.
• scripts/cacti_from_m5out.py — finds every cache in a completed run's m5out/config.json and runs Cacti on each one externally, as a stand-in for the eventual in-simulation integration.
• scripts/run_gem5_and_report.py — orchestrator: runs a gem5 config script end to end, then automatically runs whichever of the two post-hoc steps above apply (Garnet -> DSENT, caches present -> Cacti) and pulls the native power-model stats out of stats.txt, printing one consolidated report and writing it to <outdir>/power_report.txt.
• scripts/garnet-ruby-power-integration.md — the gotchas hit wiring all of the above up (ClockedObject-only power models, MathExpr quirks, DSENT path handling, ...), for adapting a different config script (topology/protocol/CPU) to the same tooling.

Usage

source activate_environment.sh

First run builds Cacti (~10s, cached in .tools/cacti/), gem5.opt for the RISCV ISA (much longer, cached via a commit stamp in .tools/, built in-place in ext/gem5/build/RISCV/), and DSENT (~10s, a patched copy cached in .tools/dsent/). Later sources are no-ops unless the respective submodule commit changed.

This exposes:

• cacti — shell function wrapping the built Cacti binary (handles a Cacti quirk where it must be run from its own directory, see the comment in the script)
• gem5 — shell function wrapping ext/gem5/build/RISCV/gem5.opt
• $CACTI_HOME, $GEM5_HOME, $GEM5_BIN, $DSENT_MODULE_DIR, $DSENT_CONFIG_DIR

One-shot: run + every applicable power/area report

python3 scripts/run_gem5_and_report.py scripts/run_minor_ruby_power.py -- \
    --cpu-type=RiscvMinorCPU --caches --ruby --network=garnet --topology=Mesh_XY --mesh-rows=1 \
    -c ext/gem5/tests/test-progs/hello/bin/riscv/linux/hello

Everything before -- is the orchestrator's own options (--outdir, --skip-dsent, --skip-cacti, ...; see --help); -- and everything after it goes to the gem5 config script untouched. This runs the simulation, then automatically runs DSENT (if the network was Garnet) and Cacti (if there are caches), and writes a combined report to m5out/power_report.txt as well as printing it.

Running the steps individually

Run a simulation with a real power model wired up in place of gem5's own se.py:

gem5 --outdir=m5out scripts/run_minor_ruby_power.py \
    --cpu-type=RiscvMinorCPU --caches --ruby --network=garnet --topology=Mesh_XY --mesh-rows=1 \
    -c ext/gem5/tests/test-progs/hello/bin/riscv/linux/hello

Then, as needed:

# Router/link power+area for a Garnet run
python3 scripts/garnet_power_from_m5out.py m5out

# Cacti estimates for every cache the run used
python3 scripts/cacti_from_m5out.py m5out

See python3 scripts/cacti_from_m5out.py --help for its options (technology node, temperature, cache type, access mode, ...).

Roadmap

• Cacti buildable standalone
• gem5 buildable standalone (RISCV/opt)
• Post-hoc Cacti analysis of a completed gem5 run's caches
• Native gem5 power model wired up for Minor CPU + Ruby caches
• Garnet NoC power/area via DSENT
• One-shot run+report orchestrator
• Call Cacti directly from gem5's C++ code so Cacti-derived numbers show up next to gem5's own stats, without a separate script step

Repo layout

• ext/ — pristine, readonly clones of external projects (git submodules)
• .tools/ — generated build state (gitignored): Cacti's build copy, gem5's build-commit stamp, DSENT's patched build copy
• cacti_runs/ — generated Cacti configs/results (gitignored)
• m5out/ — gem5 simulation output (gitignored)
• scripts/ — helper tooling