__benchmark_results__/**

"google-cloud-testutils",

"numpy",

"psutil",

"py-cpuinfo",

"pytest-benchmark",

"PyYAML",

This directory contains performance microbenchmarks for the Python Storage client library.

To run the benchmarks, use `pytest` with the `--benchmark-json` flag to specify an output file for the results.

Example:

pytest --benchmark-json=output.json -vv -s tests/perf/microbenchmarks/reads/test_reads.py

To run a single test, append `::` followed by the test name to the file path.

Example:

pytest --benchmark-json=output.json -vv -s tests/perf/microbenchmarks/reads/test_reads.py::test_downloads_single_proc_single_coro

The benchmarks are configured using `config.yaml` files located in the respective subdirectories (e.g., `reads/config.yaml`).

You can override the buckets used in the benchmarks by setting environment variables. Please refer to the specific benchmark implementation for the environment variable names.

The benchmarks produce a JSON file with the results. This file can be converted to a CSV file for easier analysis in spreadsheets using the provided `json_to_csv.py` script.

Example:

python3 tests/perf/microbenchmarks/json_to_csv.py output.json

from typing import Any, List

import statistics

import io

import os

def publish_benchmark_extra_info(

benchmark: Any,

params: Any,

benchmark_group: str = "read",

true_times: List[float] = [],

) -> None:

"""

benchmark.extra_info["num_files"] = params.num_files

benchmark.extra_info["file_size"] = params.file_size_bytes

benchmark.extra_info["chunk_size"] = params.chunk_size_bytes

if benchmark_group == "write":

benchmark.extra_info["pattern"] = "seq"

else:

benchmark.extra_info["pattern"] = params.pattern

benchmark.extra_info["coros"] = params.num_coros

benchmark.extra_info["rounds"] = params.rounds

benchmark.extra_info["bucket_name"] = params.bucket_name

benchmark.extra_info["bucket_type"] = params.bucket_type

benchmark.extra_info["processes"] = params.num_processes

benchmark.group = benchmark_group

object_size = params.file_size_bytes

num_files = params.num_files

total_uploaded_mib = (object_size / (1024 * 1024) * num_files)

min_throughput = total_uploaded_mib / benchmark.stats["max"]

max_throughput = total_uploaded_mib / benchmark.stats["min"]

mean_throughput = total_uploaded_mib / benchmark.stats["mean"]

median_throughput = total_uploaded_mib / benchmark.stats["median"]

benchmark.extra_info["throughput_MiB_s_min"] = min_throughput

benchmark.extra_info["throughput_MiB_s_max"] = max_throughput

benchmark.extra_info["throughput_MiB_s_mean"] = mean_throughput

benchmark.extra_info["throughput_MiB_s_median"] = median_throughput

print("\nThroughput Statistics (MiB/s):")

print(f" Min: {min_throughput:.2f} (from max time)")

print(f" Max: {max_throughput:.2f} (from min time)")

print(f" Mean: {mean_throughput:.2f} (approx, from mean time)")

print(f" Median: {median_throughput:.2f} (approx, from median time)")

if true_times:

throughputs = [total_uploaded_mib / t for t in true_times]

true_min_throughput = min(throughputs)

true_max_throughput = max(throughputs)

true_mean_throughput = statistics.mean(throughputs)

true_median_throughput = statistics.median(throughputs)

benchmark.extra_info["true_throughput_MiB_s_min"] = true_min_throughput

benchmark.extra_info["true_throughput_MiB_s_max"] = true_max_throughput

benchmark.extra_info["true_throughput_MiB_s_mean"] = true_mean_throughput

benchmark.extra_info["true_throughput_MiB_s_median"] = true_median_throughput

print("\nThroughput Statistics from true_times (MiB/s):")

print(f" Min: {true_min_throughput:.2f}")

print(f" Max: {true_max_throughput:.2f}")

print(f" Mean: {true_mean_throughput:.2f}")

print(f" Median: {true_median_throughput:.2f}")

# Get benchmark name, rounds, and iterations

name = benchmark.name

rounds = benchmark.stats['rounds']

iterations = benchmark.stats['iterations']

# Header for throughput table

header = "\n\n" + "-" * 125 + "\n"

header += "Throughput Benchmark (MiB/s)\n"

header += "-" * 125 + "\n"

header += f"{'Name':<50} {'Min':>10} {'Max':>10} {'Mean':>10} {'StdDev':>10} {'Median':>10} {'Rounds':>8} {'Iterations':>12}\n"

header += "-" * 125

# Data row for throughput table

# The table headers (Min, Max) refer to the throughput values.

row = f"{name:<50} {min_throughput:>10.4f} {max_throughput:>10.4f} {mean_throughput:>10.4f} {'N/A':>10} {median_throughput:>10.4f} {rounds:>8} {iterations:>12}"

print(header)

print(row)

print("-" * 125)

class RandomBytesIO(io.RawIOBase):

"""

# 10 GiB default safety cap

DEFAULT_CAP = 10 * 1024 * 1024 * 1024

def __init__(self, size, max_size=DEFAULT_CAP):

"""

if size is None:

raise ValueError("Size must be defined (cannot be infinite).")

if size > max_size:

raise ValueError(f"Requested size {size} exceeds the maximum limit of {max_size} bytes (10 GiB).")

self._size = size

self._pos = 0

def read(self, n=-1):

# 1. Handle "read all" (n=-1)

if n is None or n < 0:

n = self._size - self._pos

# 2. Handle EOF (End of File)

if self._pos >= self._size:

return b""

# 3. Clamp read amount to remaining size

# This ensures we stop exactly at `size` bytes.

n = min(n, self._size - self._pos)

# 4. Generate data

data = os.urandom(n)

self._pos += len(data)

return data

def readable(self):

return True

def seekable(self):

return True

def tell(self):

return self._pos

def seek(self, offset, whence=io.SEEK_SET):

if whence == io.SEEK_SET:

new_pos = offset

elif whence == io.SEEK_CUR:

new_pos = self._pos + offset

elif whence == io.SEEK_END:

new_pos = self._size + offset

else:

raise ValueError(f"Invalid whence: {whence}")

# Clamp position to valid range [0, size]

self._pos = max(0, min(new_pos, self._size))

return self._pos

import contextlib

from typing import Any

from tests.perf.microbenchmarks.resource_monitor import ResourceMonitor

import pytest

from tests.system._helpers import delete_blob

import asyncio

import multiprocessing

import os

import uuid

from google.cloud import storage

from google.cloud.storage._experimental.asyncio.async_appendable_object_writer import (

AsyncAppendableObjectWriter,

)

from google.cloud.storage._experimental.asyncio.async_grpc_client import AsyncGrpcClient

_OBJECT_NAME_PREFIX = "micro-benchmark"

def blobs_to_delete():

blobs_to_delete = []

yield blobs_to_delete

for blob in blobs_to_delete:

delete_blob(blob)

def storage_client():

from google.cloud.storage import Client

client = Client()

with contextlib.closing(client):

yield client

def monitor():

"""

return ResourceMonitor

def publish_resource_metrics(benchmark: Any, monitor: ResourceMonitor) -> None:

"""

benchmark.extra_info.update(

{

"cpu_max_global": f"{monitor.max_cpu:.2f}",

"mem_max": f"{monitor.max_mem:.2f}",

"net_throughput_mb_s": f"{monitor.throughput_mb_s:.2f}",

"vcpus": monitor.vcpus,

}

)

async def upload_appendable_object(bucket_name, object_name, object_size, chunk_size):

# flush interval set to little over 1GiB to minimize number of flushes.

# this method is to write "appendable" objects which will be used for

# benchmarking reads, hence not concerned performance of writes here.

writer = AsyncAppendableObjectWriter(

AsyncGrpcClient().grpc_client, bucket_name, object_name, writer_options={"FLUSH_INTERVAL_BYTES": 1026 * 1024 ** 2}

)

await writer.open()

uploaded_bytes = 0

while uploaded_bytes < object_size:

bytes_to_upload = min(chunk_size, object_size - uploaded_bytes)

await writer.append(os.urandom(bytes_to_upload))

uploaded_bytes += bytes_to_upload

object_metdata = await writer.close(finalize_on_close=True)

assert object_metdata.size == uploaded_bytes

return uploaded_bytes

def upload_simple_object(bucket_name, object_name, object_size, chunk_size):

storage_client = storage.Client()

bucket = storage_client.bucket(bucket_name)

blob = bucket.blob(object_name)

blob.chunk_size = chunk_size

data = os.urandom(object_size)

blob.upload_from_string(data)

return object_size

def _upload_worker(args):

bucket_name, object_name, object_size, chunk_size, bucket_type = args

if bucket_type == "zonal":

uploaded_bytes = asyncio.run(

upload_appendable_object(bucket_name, object_name, object_size, chunk_size)

)

else:

uploaded_bytes = upload_simple_object(bucket_name, object_name, object_size, chunk_size)

return object_name, uploaded_bytes

def _create_files(num_files, bucket_name, bucket_type, object_size, chunk_size=1024 * 1024 * 1024):

"""

object_names = [

f"{_OBJECT_NAME_PREFIX}-{uuid.uuid4().hex[:5]}" for _ in range(num_files)

]

args_list = [

(bucket_name, object_names[i], object_size, chunk_size, bucket_type)

for i in range(num_files)

]

ctx = multiprocessing.get_context("spawn")

with ctx.Pool() as pool:

results = pool.map(_upload_worker, args_list)

total_uploaded_bytes = sum(r[1] for r in results)

assert total_uploaded_bytes == object_size * num_files

return [r[0] for r in results]

def workload_params(request):

params = request.param

files_names = _create_files(

params.num_files,

params.bucket_name,

params.bucket_type,

params.file_size_bytes,

)

return params, files_names