ml_cpu_prism / README.md
David1515's picture
README: how to combine with GitHub code
a52dd85 verified
|
Raw History Blame Contribute Delete
8.92 kB
---
license: other
pretty_name: ML CPU Performance Modeling (PRISM) - gem5 traces, features and experiments
tags:
- computer-architecture
- gem5
- cpu-performance
- cpi
- simulation
size_categories:
- 100B<n<1T
---
# ml_cpu_prism
Data and code for research on machine-learning-based CPU performance (CPI) modeling with
multi-stage CPI stacks (MSCS), trained and validated against the gem5 O3 cycle-level simulator.
Everything is stored as `tar.zst` shards. To extract one:
```bash
hf download David1515/ml_cpu_prism --repo-type dataset \
--include "micro_exp/dse_exp/traces/602.gcc_s.tar.zst" --local-dir ./data
zstd -dc data/micro_exp/dse_exp/traces/602.gcc_s.tar.zst | tar -xf -
```
Each shard unpacks to the same relative path it had in the original working tree
(`micro_exp/...` shards unpack relative to `micro_exp/`; `micro_exp/dse_exp/traces/*` relative to
`micro_exp/dse_exp/traces/`, and so on).
## Using it with the code on GitHub
The code is in [github.com/davidli1515/prism_1111](https://github.com/davidli1515/prism_1111)
under `gem5-TAO/`. Its `fetch_data.sh` downloads shards from this dataset and unpacks each one
into the right place, which recreates the original working tree:
```bash
git clone https://github.com/davidli1515/prism_1111.git
cd prism_1111/gem5-TAO
./fetch_data.sh # everything, ~130 GB
./fetch_data.sh "micro_exp/exp1.tar.zst" "train_data/*" # or selected shards
```
Shards under `micro_exp/` unpack into the matching directory of `gem5-TAO/micro_exp/`;
`train_data/` and `micro_ckpt/` unpack into a separate data directory (default `data/`). See
`gem5-TAO/DATA_ON_HUGGINGFACE.md` on GitHub for the full mapping.
## Layout
| Path | Contents |
|---|---|
| `code/gem5-TAO-src.tar.zst` | Modified gem5 source tree (MSCS accounting in `src/cpu/o3/mscs.{cc,hh}`, trace hooks in O3/Minor/Simple CPUs and caches), data-generation scripts, external cache/branch-predictor simulator, Concorde analytical model |
| `micro_exp/dse_exp/traces/<benchmark>.tar.zst` | Per-slice gem5 O3 pipeline traces (`slice_N/o3trace.txt`) plus cache/TAGE warm-up checkpoints, ~400 slices per benchmark |
| `micro_exp/dse_exp/precompute/*.tar.zst` | Precomputed feature arrays (`dse_precomputed.npy`) per benchmark and per L1 size (`l1_16` ... `l1_256`) |
| `micro_exp/dse_exp/_rest.tar.zst` | Design-space-exploration scripts and results |
| `micro_exp/<experiment>.tar.zst` | Training/evaluation scripts, data, logs, results and model checkpoints for each experiment |
| `train_data/micro_traindata_packed.tar.zst` | Packed per-benchmark training arrays used by the CPI models: `concord/{benchmark}.npy` (N x 3874: 3873 Concorde features + CPI label) and `gem/{benchmark}.npy` (N x 4697: Concorde features + 808 pressure features + 16-dim MSCS target) |
| `train_data/micro_traindata_packed_all.tar.zst` | Extended packed training arrays (`gem/`) |
| `micro_ckpt/<benchmark>.tar.zst` | Per-SimPoint gem5 run directories (`<tick>/`): `simpletrace.txt` instruction traces, `stats.txt`, `config.ini/json`, logs. The checkpoints themselves (`cpt.<tick>/`, containing SPEC memory images) are **not** included, see below |
Scripts in `micro_exp` reference these arrays at `/scratch2/wl730/micro_traindata_packed{,_all}`;
extract them anywhere and update that path.
## Shards
| Shard | Size |
|---|---|
| `code/gem5-TAO-src.tar.zst` | 14.5 MB |
| `micro_ckpt/508.namd_r.tar.zst` | 0.74 GB |
| `micro_ckpt/511.povray_r.tar.zst` | 0.0 MB |
| `micro_ckpt/526.blender_r.tar.zst` | 0.74 GB |
| `micro_ckpt/600.perlbench_s.tar.zst` | 0.25 GB |
| `micro_ckpt/602.gcc_s.tar.zst` | 0.37 GB |
| `micro_ckpt/603.bwaves_s.tar.zst` | 0.0 MB |
| `micro_ckpt/605.mcf_s.tar.zst` | 0.22 GB |
| `micro_ckpt/607.cactuBSSN_s.tar.zst` | 0.71 GB |
| `micro_ckpt/619.lbm_s.tar.zst` | 0.52 GB |
| `micro_ckpt/620.omnetpp_s.tar.zst` | 0.34 GB |
| `micro_ckpt/621.wrf_s.tar.zst` | 0.21 GB |
| `micro_ckpt/623.xalancbmk_s.tar.zst` | 0.78 GB |
| `micro_ckpt/631.deepsjeng_s.tar.zst` | 0.90 GB |
| `micro_ckpt/638.imagick_s.tar.zst` | 0.74 GB |
| `micro_ckpt/641.leela_s.tar.zst` | 0.85 GB |
| `micro_ckpt/644.nab_s.tar.zst` | 0.18 GB |
| `micro_ckpt/648.exchange2_s.tar.zst` | 0.27 GB |
| `micro_ckpt/654.roms_s.tar.zst` | 0.19 GB |
| `micro_ckpt/657.xz_s.tar.zst` | 0.24 GB |
| `micro_ckpt/_files.tar.zst` | 0.0 MB |
| `micro_exp/_files.tar.zst` | 0.0 MB |
| `micro_exp/bbv.tar.zst` | 3.1 MB |
| `micro_exp/dse_exp/_rest.tar.zst` | 64.5 MB |
| `micro_exp/dse_exp/precompute/508.namd_r.tar.zst` | 9.6 MB |
| `micro_exp/dse_exp/precompute/526.blender_r.tar.zst` | 14.8 MB |
| `micro_exp/dse_exp/precompute/600.perlbench_s.tar.zst` | 21.3 MB |
| `micro_exp/dse_exp/precompute/602.gcc_s.tar.zst` | 15.1 MB |
| `micro_exp/dse_exp/precompute/605.mcf_s.tar.zst` | 22.3 MB |
| `micro_exp/dse_exp/precompute/607.cactuBSSN_s.tar.zst` | 16.6 MB |
| `micro_exp/dse_exp/precompute/619.lbm_s.tar.zst` | 39.6 MB |
| `micro_exp/dse_exp/precompute/620.omnetpp_s.tar.zst` | 54.2 MB |
| `micro_exp/dse_exp/precompute/621.wrf_s.tar.zst` | 26.8 MB |
| `micro_exp/dse_exp/precompute/623.xalancbmk_s.tar.zst` | 17.3 MB |
| `micro_exp/dse_exp/precompute/631.deepsjeng_s.tar.zst` | 32.1 MB |
| `micro_exp/dse_exp/precompute/638.imagick_s.tar.zst` | 14.3 MB |
| `micro_exp/dse_exp/precompute/641.leela_s.tar.zst` | 27.8 MB |
| `micro_exp/dse_exp/precompute/644.nab_s.tar.zst` | 12.6 MB |
| `micro_exp/dse_exp/precompute/648.exchange2_s.tar.zst` | 40.8 MB |
| `micro_exp/dse_exp/precompute/654.roms_s.tar.zst` | 13.8 MB |
| `micro_exp/dse_exp/precompute/657.xz_s.tar.zst` | 17.8 MB |
| `micro_exp/dse_exp/precompute/_files.tar.zst` | 0.0 MB |
| `micro_exp/dse_exp/precompute/l1_128.tar.zst` | 0.40 GB |
| `micro_exp/dse_exp/precompute/l1_16.tar.zst` | 0.43 GB |
| `micro_exp/dse_exp/precompute/l1_256.tar.zst` | 0.40 GB |
| `micro_exp/dse_exp/precompute/l1_32.tar.zst` | 0.41 GB |
| `micro_exp/dse_exp/precompute/l1_64.tar.zst` | 0.41 GB |
| `micro_exp/dse_exp/traces/508.namd_r.tar.zst` | 1.00 GB |
| `micro_exp/dse_exp/traces/526.blender_r.tar.zst` | 1.16 GB |
| `micro_exp/dse_exp/traces/600.perlbench_s.tar.zst` | 1.28 GB |
| `micro_exp/dse_exp/traces/602.gcc_s.tar.zst` | 1.01 GB |
| `micro_exp/dse_exp/traces/605.mcf_s.tar.zst` | 1.13 GB |
| `micro_exp/dse_exp/traces/607.cactuBSSN_s.tar.zst` | 1.09 GB |
| `micro_exp/dse_exp/traces/619.lbm_s.tar.zst` | 0.97 GB |
| `micro_exp/dse_exp/traces/620.omnetpp_s.tar.zst` | 1.14 GB |
| `micro_exp/dse_exp/traces/621.wrf_s.tar.zst` | 0.90 GB |
| `micro_exp/dse_exp/traces/623.xalancbmk_s.tar.zst` | 1.14 GB |
| `micro_exp/dse_exp/traces/631.deepsjeng_s.tar.zst` | 1.30 GB |
| `micro_exp/dse_exp/traces/638.imagick_s.tar.zst` | 1.16 GB |
| `micro_exp/dse_exp/traces/641.leela_s.tar.zst` | 1.22 GB |
| `micro_exp/dse_exp/traces/644.nab_s.tar.zst` | 0.97 GB |
| `micro_exp/dse_exp/traces/648.exchange2_s.tar.zst` | 1.28 GB |
| `micro_exp/dse_exp/traces/654.roms_s.tar.zst` | 0.89 GB |
| `micro_exp/dse_exp/traces/657.xz_s.tar.zst` | 1.15 GB |
| `micro_exp/dse_v2.tar.zst` | 3.81 GB |
| `micro_exp/exp1.tar.zst` | 2.52 GB |
| `micro_exp/exp2/exp2.tar.zst.part00` | 21.47 GB |
| `micro_exp/exp2/exp2.tar.zst.part01` | 21.47 GB |
| `micro_exp/exp2/exp2.tar.zst.part02` | 21.47 GB |
| `micro_exp/exp2/exp2.tar.zst.part03` | 18.13 GB |
| `micro_exp/exp4-gradient-conflict.tar.zst` | 0.4 MB |
| `micro_exp/exp4.tar.zst` | 10.18 GB |
| `micro_exp/exp5.tar.zst` | 0.56 GB |
| `micro_exp/exp5_work_410.tar.zst` | 0.11 GB |
| `micro_exp/exp5_work_528.tar.zst` | 0.17 GB |
| `micro_exp/lmono_reuslts.tar.zst` | 2.9 MB |
| `micro_exp/lmono_sweep.tar.zst` | 0.0 MB |
| `micro_exp/mcpat.tar.zst` | 3.7 MB |
| `micro_exp/new_bench.tar.zst` | 29.8 MB |
| `micro_exp/pretrain.tar.zst` | 0.0 MB |
| `micro_exp/window_ids.tar.zst` | 0.6 MB |
| `train_data/micro_traindata_packed.tar.zst` | 0.32 GB |
| `train_data/micro_traindata_packed_all.tar.zst` | 0.58 GB |
| **Total (83 files)** | **130.4 GB** |
`micro_exp/exp2` (mostly model checkpoints, 82.6 GB) is split into 20 GB parts because of the per-file size limit. Reassemble with:
```bash
cat micro_exp/exp2/exp2.tar.zst.part* | zstd -dc | tar -xf -
```
## Provenance and licensing
- Traces were produced by simulating SPEC CPU2017 workloads in gem5. **No SPEC CPU2017
binaries, source code or input files are included**; only simulator output is distributed.
Reproducing the traces requires your own licensed copy of SPEC CPU2017.
- The gem5 SimPoint checkpoints (whose memory images contain SPEC binaries and inputs) and the
SPEC run directories needed to restart from them are kept in a separate private dataset,
shared only with holders of a SPEC CPU2017 license. Contact the author for access.
- The gem5 source is distributed under gem5's original BSD-style license (see `LICENSE`/`COPYING`
inside `code/gem5-TAO-src.tar.zst`).
- `micro_exp/new_bench` contains a bundle of open-source benchmarks under their own licenses.