Dataset Benchmarking¶

This notebook is a behind-the-scenes benchmarking notebook, mainly for use by developers. The recommended way for users to interact with the dataset is via the Measurement object and its associated context manager. See the corresponding notebook for a comprehensive toturial on how to use those.

In [1]:

%matplotlib inline
from pathlib import Path

import numpy as np

import qcodes as qc
from qcodes.dataset import (
    ParamSpec,
    initialise_or_create_database_at,
    load_or_create_experiment,
    new_data_set,
)

In [2]:

qc.config.core.db_location

Out[2]:

'~/experiments.db'

In [3]:

initialise_or_create_database_at(
    Path.cwd().parent / "example_output" / "benchmarking.db"
)

Setup¶

In [4]:

exp = load_or_create_experiment("benchmarking", sample_name="the sample is a lie")
exp

Out[4]:

benchmarking#the sample is a lie#1@/home/runner/work/Qcodes/Qcodes/docs/examples/example_output/benchmarking.db
---------------------------------------------------------------------------------------------------------------

Now we can create a dataset. Note two things:

- if we don't specfiy a exp_id, but we have an experiment in the experiment container the dataset will go into that one.
- dataset can be created from the experiment object

In [5]:

dataSet = new_data_set("benchmark_data", exp_id=exp.exp_id)
exp

Out[5]:

benchmarking#the sample is a lie#1@/home/runner/work/Qcodes/Qcodes/docs/examples/example_output/benchmarking.db
---------------------------------------------------------------------------------------------------------------
1-benchmark_data-1-None-0

In this benchmark we will assueme that we are doing a 2D loop and investigate the performance implications of writing to the dataset

In [6]:

x_shape = 100
y_shape = 100

Baseline: Generate data¶

In [7]:

%%time
for x in range(x_shape):
    for y in range(y_shape):
        z = np.random.random_sample(1)

CPU times: user 18.6 ms, sys: 0 ns, total: 18.6 ms
Wall time: 18.3 ms

and store in memory

In [8]:

x_data = np.zeros((x_shape, y_shape))
y_data = np.zeros((x_shape, y_shape))
z_data = np.zeros((x_shape, y_shape))

In [9]:

%%time
for x in range(x_shape):
    for y in range(y_shape):
        x_data[x, y] = x
        y_data[x, y] = y
        z_data[x, y] = np.random.random_sample()

CPU times: user 15.1 ms, sys: 2 μs, total: 15.1 ms
Wall time: 14.9 ms

Add to dataset inside double loop¶

In [10]:

double_dataset = new_data_set(
    "doubledata",
    exp_id=exp.exp_id,
    specs=[
        ParamSpec("x", "numeric"),
        ParamSpec("y", "numeric"),
        ParamSpec("z", "numeric"),
    ],
)
double_dataset.mark_started()

Note that this is so slow that we are only doing a 10th of the computation

In [11]:

%%time
for x in range(x_shape // 10):
    for y in range(y_shape):
        double_dataset.add_results([{"x": x, "y": y, "z": np.random.random_sample()}])

CPU times: user 189 ms, sys: 74.4 ms, total: 263 ms
Wall time: 702 ms

Add the data in outer loop and store as np array¶

In [12]:

single_dataset = new_data_set(
    "singledata",
    exp_id=exp.exp_id,
    specs=[ParamSpec("x", "array"), ParamSpec("y", "array"), ParamSpec("z", "array")],
)
single_dataset.mark_started()
x_data = np.zeros(y_shape)
y_data = np.zeros(y_shape)
z_data = np.zeros(y_shape)

In [13]:

%%time
for x in range(x_shape):
    for y in range(y_shape):
        x_data[y] = x
        y_data[y] = y
        z_data[y] = np.random.random_sample(1)
    single_dataset.add_results([{"x": x_data, "y": y_data, "z": z_data}])

CPU times: user 43.5 ms, sys: 5.03 ms, total: 48.5 ms
Wall time: 64.1 ms

<timed exec>:5: DeprecationWarning: Conversion of an array with ndim > 0 to a scalar is deprecated, and will error in future. Ensure you extract a single element from your array before performing this operation. (Deprecated NumPy 1.25.)

Save once after loop¶

In [14]:

zero_dataset = new_data_set(
    "zerodata",
    exp_id=exp.exp_id,
    specs=[ParamSpec("x", "array"), ParamSpec("y", "array"), ParamSpec("z", "array")],
)
zero_dataset.mark_started()
x_data = np.zeros((x_shape, y_shape))
y_data = np.zeros((x_shape, y_shape))
z_data = np.zeros((x_shape, y_shape))

In [15]:

%%time
for x in range(x_shape):
    for y in range(y_shape):
        x_data[x, y] = x
        y_data[x, y] = y
        z_data[x, y] = np.random.random_sample(1)
zero_dataset.add_results([{"x": x_data, "y": y_data, "z": z_data}])

CPU times: user 18.7 ms, sys: 897 μs, total: 19.6 ms
Wall time: 19.8 ms

<timed exec>:5: DeprecationWarning: Conversion of an array with ndim > 0 to a scalar is deprecated, and will error in future. Ensure you extract a single element from your array before performing this operation. (Deprecated NumPy 1.25.)

Array parameter¶

In [16]:

array1D_dataset = new_data_set(
    "array1Ddata",
    exp_id=exp.exp_id,
    specs=[ParamSpec("x", "array"), ParamSpec("y", "array"), ParamSpec("z", "array")],
)
array1D_dataset.mark_started()
y_setpoints = np.arange(y_shape)

In [17]:

%%timeit
for x in range(x_shape):
    x_data[x, :] = x
    array1D_dataset.add_results(
        [{"x": x_data[x, :], "y": y_setpoints, "z": np.random.random_sample(y_shape)}]
    )

43.6 ms ± 1.89 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [18]:

x_data = np.zeros((x_shape, y_shape))
y_data = np.zeros((x_shape, y_shape))
z_data = np.zeros((x_shape, y_shape))
y_setpoints = np.arange(y_shape)

In [19]:

array0D_dataset = new_data_set(
    "array0Ddata",
    exp_id=exp.exp_id,
    specs=[ParamSpec("x", "array"), ParamSpec("y", "array"), ParamSpec("z", "array")],
)
array0D_dataset.mark_started()

In [20]:

%%timeit
for x in range(x_shape):
    x_data[x, :] = x
    y_data[x, :] = y_setpoints
    z_data[x, :] = np.random.random_sample(y_shape)
array0D_dataset.add_results([{"x": x_data, "y": y_data, "z": z_data}])

1.86 ms ± 29.1 μs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Insert many¶

In [21]:

data = []
for i in range(100):
    for j in range(100):
        data.append({"x": i, "y": j, "z": np.random.random_sample()})

In [22]:

many_Data = new_data_set(
    "many_data",
    exp_id=exp.exp_id,
    specs=[
        ParamSpec("x", "numeric"),
        ParamSpec("y", "numeric"),
        ParamSpec("z", "numeric"),
    ],
)
many_Data.mark_started()

In [23]:

%%timeit
many_Data.add_results(data)

17.8 ms ± 544 μs per loop (mean ± std. dev. of 7 runs, 100 loops each)