Skip to content

Latest commit

 

History

History
47 lines (37 loc) · 1.58 KB

run_experiments.md

File metadata and controls

47 lines (37 loc) · 1.58 KB
Raw

Run experiments

Accelerators:

  • GPU: Nvidia GeForce GTX 460 SE
  • FPGA: Xilinx Virtex Ultrascale+ VCU118

Realm VM Boot

Realm VM:

# pwd is ./scripts
# XXX: core 3 is reserved for tracing

core=0x4
root=../assets/snapshots

nice -n -20 taskset $core ./$root/lkvm run --realm  --disable-sve -c 1 -m 1000 \
    -k $root/Image-cca -i $root/rootfs.realm.cpio --9p "/,host0" -p "fvp_escape_loop fvp_escape_on ip=off"

Build

  • ./scripts/install.sh (see ./doc/ on how to build).

GPU

  • Benchmarks: ./src/gpu_driver/rodinia-bench/cuda/setup.sh
  • Pre Benchmark Setup:
    • Load ACAI Helper Program: ./scripts/fvp/setup_devmem_intercept_realm.sh
    • Load GPU Driver: ./scripts/fvp/setup_fvp_gpu.sh

FPGA

  • Benchmarks: ./src/benchmarking/fpga/setup.sh
  • Pre Benchmark Setup:
    • Load ACAI Helper Program: ./scripts/fvp/setup_devmem_intercept_realm.sh
    • Load FPGA Driver: ./scripts/fvp/setup_fvp_fpga.sh

Gathering Data

  • We implement a custom tracer plugin to gather executed instructions in the FVP.
  • The setup scripts already instrument the FVP to use this plugin.
  • Upon benchmark completion, the tracer plugin creates data dumps in /tmp/arm_* on the x86 host system running the FVP.
  • The benchmarks are instrumented to use marker instructions to signalize special events.
    • Grep for (CCA_ marcors in the benchmarks)
    • For instance: #define CCA_MARKER(marker) __asm__ volatile("MOV XZR, " STR(marker))
    • The MOV instruction is used to store a marker value into the zero register (XZR), which is a NOP and is detected by the tracer plugin as event of special interest.