SDRBench — SCALE-LETKF

Weather simulation from the SCALE-RM model with LETKF data assimilation (RIKEN): 12 variables on 98 x 1200 x 1200 (z, y, x), single precision, little-endian.

Raw binary arrays, not an Arrow/Parquet dataset. datasets.load_dataset("sdrbench/scale-letkf") returns the table of files (files.jsonl: path, dtype, shape, sha256). Read the arrays with the sdrbench package or hf_hub_download + numpy as shown below.

This repository is an unmodified mirror of the SCALE-LETKF data of SDRBench, the Scientific Data Reduction Benchmark. The originals are hosted by Argonne National Laboratory on Globus; every archive was unpacked and its files uploaded byte-for-byte (sha256 verified) (revision 18665ddfb9). metadata/ holds the original SDRBench property and template files. Kept in sync automatically by szcompressor/sdrbench.

• Data provider: RIKEN; contact Guo-Yuan Lien (guoyuan.lien@gmail.com)
• Size: 13.5 GB in 2 variant(s)
• Format: raw little-endian binary, C order (slowest dimension first); dtype and shape per field below

Usage

pip install "sdrbench[sz3]"     # sdrbench + pysz (SZ3)

import numpy as np
import sdrbench
from pysz import sz, szConfig, szErrorBoundMode

ds = sdrbench.dataset("scale-letkf")                # default variant "original"
print(ds.fields)
x = ds["PRES"]                     # numpy array, dtype <f4, shape (98, 1200, 1200)

# compress with SZ3 (pysz) at a 1e-3 value-range-relative error bound
conf = szConfig()
conf.errorBoundMode = szErrorBoundMode.REL
conf.relErrorBound = 1e-3
compressed, ratio = sz.compress(np.ascontiguousarray(x), conf)
y, _ = sz.decompress(compressed, x.dtype.type, x.shape)
max_err, psnr, nrmse = sz.verify(np.asarray(x), y)
print(f"ratio {ratio:.1f}x, PSNR {psnr:.1f} dB, max error {max_err:.3g}")

Files are fetched on first use (in parallel with ds.download("data/")) and cached, pinned to the revision of the installed sdrbench release; if Hugging Face is unreachable the package falls back to the original archive on Globus. Existing local SDRBench copies can be used with sdrbench.dataset("scale-letkf", root="/path").

Without the package, with huggingface_hub and numpy only:

from huggingface_hub import hf_hub_download
import numpy as np
p = hf_hub_download("sdrbench/scale-letkf", "98x1200x1200/PRES-98x1200x1200.f32", repo_type="dataset", revision="18665ddfb989801761218b81e665072f998a23ff")
x = np.fromfile(p, dtype="<f4").reshape((98, 1200, 1200))

Variants and fields

original (default)

sdrbench.dataset("scale-letkf", "original") — folder 98x1200x1200/ — 12 files, 6.8 GB, from SDRBENCH-SCALE-98x1200x1200.tar.gz.

All variables.

Field	dtype	Shape (C order)	File
PRES	<f4	98 x 1200 x 1200	98x1200x1200/PRES-98x1200x1200.f32
QC	<f4	98 x 1200 x 1200	98x1200x1200/QC-98x1200x1200.f32
QG	<f4	98 x 1200 x 1200	98x1200x1200/QG-98x1200x1200.f32
QI	<f4	98 x 1200 x 1200	98x1200x1200/QI-98x1200x1200.f32
QR	<f4	98 x 1200 x 1200	98x1200x1200/QR-98x1200x1200.f32
QS	<f4	98 x 1200 x 1200	98x1200x1200/QS-98x1200x1200.f32
QV	<f4	98 x 1200 x 1200	98x1200x1200/QV-98x1200x1200.f32
RH	<f4	98 x 1200 x 1200	98x1200x1200/RH-98x1200x1200.f32
T	<f4	98 x 1200 x 1200	98x1200x1200/T-98x1200x1200.f32
U	<f4	98 x 1200 x 1200	98x1200x1200/U-98x1200x1200.f32
V	<f4	98 x 1200 x 1200	98x1200x1200/V-98x1200x1200.f32
W	<f4	98 x 1200 x 1200	98x1200x1200/W-98x1200x1200.f32

log

sdrbench.dataset("scale-letkf", "log") — folder 98x1200x1200_log/ — 12 files, 6.8 GB, from SDRBENCH-SCALE-98x1200x1200_log.tar.gz.

The hydrometeor and vapour mixing ratios log10-transformed; the other variables unchanged.

Field	dtype	Shape (C order)	File
PRES	<f4	98 x 1200 x 1200	98x1200x1200_log/PRES-98x1200x1200.f32
QC_log10	<f4	98 x 1200 x 1200	98x1200x1200_log/QC-98x1200x1200.log10.f32
QG_log10	<f4	98 x 1200 x 1200	98x1200x1200_log/QG-98x1200x1200.log10.f32
QI_log10	<f4	98 x 1200 x 1200	98x1200x1200_log/QI-98x1200x1200.log10.f32
QR_log10	<f4	98 x 1200 x 1200	98x1200x1200_log/QR-98x1200x1200.log10.f32
QS_log10	<f4	98 x 1200 x 1200	98x1200x1200_log/QS-98x1200x1200.log10.f32
QV_log10	<f4	98 x 1200 x 1200	98x1200x1200_log/QV-98x1200x1200.log10.f32
RH	<f4	98 x 1200 x 1200	98x1200x1200_log/RH-98x1200x1200.f32
T	<f4	98 x 1200 x 1200	98x1200x1200_log/T-98x1200x1200.f32
U	<f4	98 x 1200 x 1200	98x1200x1200_log/U-98x1200x1200.f32
V	<f4	98 x 1200 x 1200	98x1200x1200_log/V-98x1200x1200.f32
W	<f4	98 x 1200 x 1200	98x1200x1200_log/W-98x1200x1200.f32

Citation

Please cite SDRBench; sdrbench cite scale-letkf prints it:

@inproceedings{zhao2020sdrbench,
  title     = {{SDRBench}: Scientific Data Reduction Benchmark for Lossy Compressors},
  author    = {Zhao, Kai and Di, Sheng and Liang, Xin and Li, Sihuan and Tao, Dingwen and Chen, Zizhong and Cappello, Franck},
  booktitle = {2020 IEEE International Conference on Big Data (Big Data)},
  pages     = {2716--2724},
  year      = {2020}
}

License and terms

The data are distributed by the SDRBench team for research use under the terms of the original data providers listed above (see the SDRBench website); please acknowledge them as requested. Rights remain with the data providers.