This is a research project out of the Boston University Computer Science Department
Computers used to be large, expensive and centrally owned. Operating systems were designed to multiplex hardware across multiple applications and users. Presently, Cloud Computing allows users to rent entire virtual machines to run their own OS and applications. In this context, the hypervisor multiplexes hardware across multiple applications and users. We can build more efficient software by constructing custom, application-specific operating systems rather than deploying general purpose operating systems within virtual machines.
Using EbbRT, we are able to achieve ~2x better throughput at SLA for memcached.
A port of the node.js javascript environment obtained ~4% improvement on pure compute workloads (no I/O).
EbbRT is designed first and foremost for performance and maintainability. A high performance system that cannot be maintained will soon fall behind. This is one of the primary pitfalls of customized systems. EbbRT adopts several techniques in order to achieve this:
-
EbbRT is comprised of a set of components that developers can extend, replace or discard in order to construct and deploy a particular application. This enables a much greater degree of customization than existing general purpose systems while promoting the reuse of non-performance-critical components
-
EbbRT components run in a light-weight event-driven environment. This reduces the runtime complexity yet provides enough flexibility for a wide range of applications.
-
EbbRT library operating systems can run within virtual machines on unmodified hypervisors. This allows us to deploy EbbRT applications on commodity clouds.
-
EbbRT library operating systems run alongside general purpose operating systems. This allows functionality to be offloaded for compatibility, reducing the maintenance burden by avoiding the construction of new software.
-
EbbRT uses many modern and high-level programming techniques not typically found in operating systems software. This was chosen deliberately to reduce the complexity of the software.
EbbRT is comprised of a set of an x86_64 library OS and toolchain as well as a Linux userspace library. Both runtimes are written predominately in C++11. The native library OS is deployed along with a modified GNU toolchain (gcc, binutils, libstdc++) and newlib (libc) that provide an x86_64-ebbrt target. Application code targeting the native library OS is compiled with this toolchain and the resulting binary is a bootable ELF linked with the library OS. We provide C and C++ standard library implementations which make it straightforward to use many third party software libraries.
Checkout the repository submodules to pull in the build requirements.
git submodule update --init
Set the EBBRT_SRCDIR environment variable to point to EbbRT source
directory. This variable is require to build an EbbRT application.
export EBBRT_SRCDIR=$PWD
Step to the EbbRT toolchain submodule directory and run make. By default,
this will download the latest toolchain binaries. For instructions to build the toolchain from
source see the README file in the EbbRT-toolchain repository.
make -C $EBBRT_SRCDIR/baremetal/ext/EbbRT-toolchain
To build an EbbRT application run make in the application's directory. Make sure the EBBRT_SCRDIR environment variable is available and points to the EbbRT source directory.
make -j -C $EBBRT_SRCDIR/apps/helloworld
Our latest technical report (PDF)
Please email us at sesa-developers@googlegroups.com