Photon is a high-efficiency LibOS framework, based on a set of carefully selected C++ libs.

The role of LibOS is to connect user apps and the kernel. Following the principle of Least Astonishment, we designed Photon's API to be as consistent as possible with glibc and POSIX semantics. This flattens the learning curve for lib users and brings convenience when migrating legacy codebases.

Photon's runtime is driven by a coroutine lib. Out tests show that it has the best I/O performance in the open source world by the year of 2022, even among different programing languages.

As to the project vision, we hope that Photon would help programs run as fast and agile as the photon particle, which exactly is the naming came from.

• Coroutine library (support multi-core)
• Async event engine, natively integrated into coroutine scheduling (support epoll or io_uring)
• Multiple I/O wrappers: psync, posix_aio, libaio, io_uring
• Multiple socket implementations: tcp (level-trigger/edge-trigger), unix-domain, zero-copy, libcurl, TLS support, etc.
• A high performance and lightweight RPC client/server
• A HTTP client/server (even faster than Nginx)
• A POSIX-like filesystem abstraction and some implementations: local fs, http fs, fuse fs, etc.
• A bunch of useful tools: io-vector manipulation, resource pool, object cache, mem allocator, callback delegator, pre-compiled logging, lockless ring buffer, etc.

While Photon has already encapsulated many mature OS functionalities, it remains keen to the latest kernel features, and prepared to wrap them into the framework. It is a real killer in the low level programing field.

Compare Photon and fio when reading an 3.5TB NVMe raw device.

Note that fio only enables 1 job (process).

Conclusion: Photon is faster than fio under this circumstance.

Compare TCP echo server performance, in descending order.

Client Mode: Streaming

Client Mode: Ping-pong

Note:

• The Streaming client is to measure echo server performance when handling high throughput. We will set up 4 client processes, and each of them will create only one connection. Send coroutine and recv coroutine are running their loops separately.
• The Ping-pong client is to measure echo server performance when handling large amounts of connections. We will set up 10 client processes, and each of them will create 100 connections. For a single connection, it has to send before recv.
• Server and client are all cloud VMs, 64Core 128GB, Intel Platinum CPU 2.70GHz. Kernel version is 5.15. The network bandwidth (unilateral) is 32Gb.
• This test was only meant to compare per-core QPS, so we limited the thread number to 1, for instance, set GOMAXPROCS=1.

Conclusion: Photon socket has the best per-core QPS.

Compare Photon and Nginx when serving static files, using Apache Bench(ab) as client.

Note that Nginx only enables 1 worker (process).

Conclusion: Photon is faster than Nginx under this circumstance.

See the simple example (on io_uring) about how to write a Photon program.

See the full test code of echo server. It also illustrates how to enable multi-core.

dnf install gcc-c++ epel-release cmake
dnf install openssl-devel libcurl-devel libaio-devel

apt install cmake
apt install libssl-dev libcurl4-openssl-dev libaio-dev

mkdir build && cd build
cmake ..
make -j

All the libs and executables will be saved in build/output.

# CentOS
dnf config-manager --set-enabled powertools
dnf install gtest-devel gmock-devel gflags-devel fuse-devel libgsasl-devel
# Ubuntu
apt install libgtest-dev libgmock-dev libgflags-dev libfuse-dev libgsasl7-dev

cmake -D BUILD_TESTING=1 -D ENABLE_SASL=1 -D ENABLE_FUSE=1 -D CMAKE_BUILD_TYPE=Debug ..
make -j
ctest

Note the examples are also built from testing. When running performance test, remember to remove CMAKE_BUILD_TYPE=Debug.

Photon was originally created from the storage team of Alibaba Cloud since 2017. It's a production ready library, and has been deployed to hundreds of thousands of hosts as the infrastructure of cloud software. We would like to make a commitment that Photon will be continuously updated, as long as those cloud software still evolve.

Some open source projects are using Photon as well, for instance:

• containerd/overlaybd The storage backend of accelerated container image, providing a layering block-level image format, designed for container, secure container and virtual machine.
• data-accelerator/photon-libtcmu A TCMU implementation, reworked from tcmu-runner, acting as a iSCSI target.

Any addition to this list is appreciated, if you have been using Photon, or just enlightened by its coroutine design.

We are building an independent website for developers to view the documents. Please stay tuned.