Blobfuse2 - A Microsoft supported Azure Storage FUSE driver

About

Blobfuse2 is an open source project developed to provide a virtual filesystem backed by the Azure Storage. It uses the libfuse open source library (fuse3) to communicate with the Linux FUSE kernel module, and implements the filesystem operations using the Azure Storage REST APIs.

Performance Benchmarking

To keep track of performance regression introduced by any commit in the main branch, we run a continuous benchmark test suite. This test suite is executed periodically and for each commit done to main branch. Suite uses fio, a industry standard storage benchmarking tool, and few custom applications to perform various tests. Results are then published in our gh-pages in this repository for a visual representaiton.

Setup

VM

All tests are performed on Standard D96ds_v5 (96 vcpus, 384 GiB memory) Azure VM running in eastus2 region. Specifications of this VM is listed here.

Storage

A Premium Blob Storage Account and a Standard Blob Storage Account in eastus2 region were used to conduct all tests. HNS was disabled on both these accounts.

Blobfuse2 configuration

Blobfuse2 is configured with block-cache for all tests. Other than Large Threads case persistance of blocks on the disk was disabled. Configuration file used in this test is available here for reference.

Tests and Results

Master test script that simulates this benchmarking test suite is located here. To execute a specific test case download the script and execute below command:

    fio_bench.sh <mount-path> <test-name>

• Script expects Blobfuse2 is already install on the system.
• Script expects the config file is present in current directory with name config.yaml
• Install fio and jq before you execute the script
• Allowed test-name are: read / write / create / list / app / rename

Below table provides latency/time and bandwidth results for various tests on respective account types. Each test has a linked section describing details of that test case.

Test Name	Premium Blob Storage Account	Standard Blob Storage Account
Create	Latency : Bandwidth	Latency : Bandwidth
Write	Latency : Bandwidth	Latency : Bandwidth
Read	Latency : Bandwidth	Latency : Bandwidth
List	Time	Time
Rename	Time	Time
Highly Parallel	Latency : Bandwidth	Latency : Bandwidth
Application	Time : Bandwidth	Time : Bandwidth
Max Out	Time : Bandwidth	Time : Bandwidth

Create

In this test fio command is used to create large number of small files in parallel. As part of the test bandwidth and latency are measured. Following cases are performed as part of this test:

    - Create 1000 files of 1M each in 10 parallel threads
    - Create 1000 files of 1M each in 100 parallel threads
    - Create 100,000 files of 1K each in 20 parallel threads

All fio config files used during these tests are located here.

Write

In this test fio command is used to run various write workflows. As part of the test bandwidth and latency are measured. Each test is run for 30 seconds and average of 3 such iteration is taken. Both bandwidth and latency are taken from fio output directly and projected in the charts. To simulate different write work-flows following cases are performed:

    - Sequential write on a 100G file
    - Sequential write on a 100G file with direct-io
    - Sequential write by 4 parallel threads on 4 different files of 100G size.
    - Sequential write by 16 parallel threads on 16 different files of 100G size.

All fio config files used during these tests are located here.

Read

In this test fio command is used to run various read workflows. As part of the test bandwidth and latency are measured. Each test is run for 30 seconds and average of 3 such iteration is taken. Both bandwidth and latency are taken from fio output directly and projected in the charts. To simulate different read work-flows following cases are performed:

    - Sequential read on a 100G file
    - Random read on a 100G file
    - Sequential read on a small 5M file
    - Random read on a small 5M file
    - Sequential read on a 100G file with direct-io
    - Random read on a 100G file with direct-io
    - Sequential read on a 100G file by 4 parallel threads
    - Sequential read on a 100G file by 16 parallel threads
    - Random read on a 100G file by 4 parallel threads

All fio config files used during these tests are located here.

List

In this test standard linux ls -U --color=never command is used to list all files on the mount path. Graphs are plotted based on total time taken by this operation. Post listing this test case also execute a delete operation to delete all files on the container and measures total time take to delete these files. For the benchmarking purpose this test is executed after Create tests so that storage container has 100K+ files.

Rename

In this test 5000 files of 1MB size each are created using a python script and then each file is renamed using os.rename() method. Total time to rename all 5000 files is measured.

Highly Parallel

In this test fio command is used to run various read/write workflows with high number of parallel threads. As part of the test bandwidth and latency are measured. Each test is run for 60 seconds and average of 3 such iteration is taken. Both bandwidth and latency are taken from fio output directly and projected in the charts. To simulate different read/write work-flows following cases are performed:

    - Sequential write of 1G file each by 112 thread
    - Sequential read of 1G file each by 128 thread
    - Sequential read of 1G file each by 128 thread with direct-io

All fio config files used during these tests are located here.

Application Test

We have observed that fio and dd commands have certain overhead of their own and they are not able to utilize our fuse solution to its full potential. With such tools you will observe tool itself consuming 100% CPU and performance of Blobfuse2 being blocked by the tool itself. To overcome this we created custom python applications to simulate sequential red/write of a large file.

These applications are located here. Read/Write application is executed once and based on the time taken to read/write given amount of data bandwidth is computed.

Max Out

Objective of this test case is to observe the max throughput that Blobfuse2 can generate. To achieve this two tests are run:

    - Create 10 20GB files
    - Read 10 20GB files

In case test 10 threads are run in parallel. Each thread is bound to a CPU core so that it gets max CPU for operation. 'dd' command is used to execute the test.

For write operation, data is read from '/dev/zero' in 16MB chunks and written to file on mounted path with 'direct io'.

For read operations, data is read from mount path in 4MB chunks and sent to '/dev/null' with 'dd' set to not report any status.

Test Scripts are located here with name "highspeed_".