1343503048 / ghost-userspace

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

ghOSt: Fast & Flexible User-Space Delegation of Linux Scheduling

ghOSt is a general-purpose delegation of scheduling policy implemented on top of the Linux kernel. The ghOSt framework provides a rich API that receives scheduling decisions for processes from userspace and actuates them as transactions. Programmers can use any language or tools to develop policies, which can be upgraded without a machine reboot. ghOSt supports policies for a range of scheduling objectives, from µs-scale latency, to throughput, to energy efficiency, and beyond, and incurs low overheads for scheduling actions. Many policies are just a few hundred lines of code. Overall, ghOSt provides a performant framework for delegation of thread scheduling policy to userspace processes that enables policy optimization, non-disruptive upgrades, and fault isolation.

The ghOSt kernel is here. You must compile and run the userspace component on the ghOSt kernel.

This is not an officially supported Google product.


The ghOSt userspace component can be compiled on Ubuntu 20.04 or newer.

1. We use the Google Bazel build system to compile the userspace components of ghOSt. Go to the Bazel Installation Guide for instructions to install Bazel on your operating system.

2. Install ghOSt dependencies:

sudo apt update
sudo apt install libnuma-dev libcap-dev libelf-dev libbfd-dev clang llvm zlib1g-dev python-is-python3

3. Compile the ghOSt userspace component. Run the following from the root of the repository:

bazel build -c opt ...

-c opt tells Bazel to build the targets with optimizations turned on. ... tells Bazel to build all targets in the BUILD file and all BUILD files in subdirectories, including the core ghOSt library, the eBPF code, the schedulers, the unit tests, the experiments, and the scripts to run the experiments, along with all of the dependencies for those targets. If you prefer to build individual targets rather than all of them to save compile time, replace ... with an individual target name, such as agent_shinjuku.

About

License:Apache License 2.0


Languages

Language:C 72.7%Language:C++ 24.6%Language:Python 1.7%Language:Starlark 0.9%