[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [RFC] Unicore Subproject Proposal

Hi Lars, all,

[cc’ing authors of Erlang on Xen, HalVM and Rump].

Thanks everyone for all of the support and useful comments. We’ve
incorporated a number of them into a new version of the document (attached
and pasted at the bottom for convenience) and for those that didn’t make
it we’re keeping track of them.

Lars, FYI, Simon also did a blog post regarding Unicore on unikernel.org

Please let us know what the next steps are.


— Felipe


Project Leads:    Simon Kuenzer      <simon.kuenzer@xxxxxxxxx>
     (co-lead)    Felipe Huici       <felipe.huici@xxxxxxxxx>
     (co-lead)    Florian Schmidt    <florian.schmidt@xxxxxxxxx>
Project Mentor:   Lars Kurth         <lars.kurth@xxxxxxxxxx>
Project Sponsors: Stefano Stabellini <sstabellini@xxxxxxxxxx>
                  Wei Liu            <wei.liu2@xxxxxxxxxx>

In recent years, several papers and projects dedicated to unikernels
have shown the immense potential for performance gains that these
have. By leveraging specialization and the use of minimalistic OSes,
unikernels are able to yield impressive numbers, including fast
instantiation times (tens of milliseconds or less), tiny memory
footprints (a few MBs or even KBs), high network throughput (10-40
Gb/s), and high consolidation (e.g., being able to run thousands of
instances on a single commodity server), not to mention a reduced
attack surface and the potential for easier certification. Unikernel
projects worthy of mention include MirageOS, ClickOS, Erlang on Xen,
OSv, HALVM, and Minicache, Rump, among others.

The fundamental drawback of unikernels is that they require that
applications be manually ported to the underlying minimalistic OS (e.g.
having to port nginx, snort, mysql or memcached to MiniOS or OSv); this
requires both expert work and often considerable amount of time. In
essence, we need to pick between either high performance
with unikernels, or no porting effort but decreased performance
and decreased efficiency with standard OS/VM images.
The goal of this proposal is to change this status quo by providing
a highly configurable unikernel code base; we call this base Unicore.

This project also aims to concentrate the various efforts currently going
on in the Xen community regarding minimalistic OSes (essentially different
variants of MiniOS). We think that splitting the community across these
variants is counter-productive and hope that Unicore will provide a common
place for all or most improvements and customizations of minimalistic
OSes. The long term goal is to replace something like MiniOS with a tool
that can automatically build such a minimalistic OS.

Unicore - The "Unikernel Core"
The high level goal of Unicore is to be able to build unikernels targeted
at specific applications without requiring the time-consuming, expert work
that building such a unikernel requires today. An additional goal (or
hope) of Unicore is that all developers interested in unikernel
development would contribute by supplying libraries rather than working on
independent projects with different code bases as it is done now. The main
idea behind Unicore is depicted in Figure 1 and consists of two basic

[Attachment: unicore-oneslider.pdf]

Figure 1. Unicore Architecture.

Library pools would contain libraries that the user of Unicore can select
from to create the unikernel. From the bottom up, library pools are
organized into (1) the architecture library tool, containing libraries
specific to a computer architecture (e.g., x86_64, ARM32 or MIPS); (2) the
platform tool, where target platforms can be Xen, KVM, bare metal (i.e. no
virtualization) and user-space Linux; and (3) the main library pool,
containing a rich set of functionality to build the unikernel from. This
last library includes drivers (both virtual such as netback/netfront and
physical such as ixgbe), filesystems, memory allocators, schedulers,
network stacks, standard libs (e.g. libc, openssl, etc.), runtimes (e.g. a
Python interpreter and debugging and profiling tools. These pools of
libraries constitute a code base for creating unikernels. As shown, a
library can be relatively large (e.g libc) or quite small (a scheduler),
which should allow for a fair amount of customization for the unikernel.
The Unicore build tool is in charge of compiling the application and the
selected libraries together to create a binary for a specific platform and
architecture (e.g., Xen on x86_64). The tool is currently inspired by
Linux’s kconfig system and consists of a set of Makefiles. It allows users
to select libraries, to configure them, and to warn them when library
dependencies are not met. In addition, the tool can also simultaneously
generate binaries for multiple platforms.
As an example, imagine a user wanting to generate a network driver domain
unikernel. In this case, we would assume the “application” to be the
netback driver. To select this application, the user would first run “make
menuconfig” from within the netback application folder. The Makefile there
would set a variable to indicate what the application is, and would
include the main Unicore Makefiles so that the unikernel can be built
(Step 1 in the figure). Using the menu-based system, the user chooses the
relevant libraries; for a Xen driver domain this would include a physical
network driver, the netback driver, the libxenplat library and a library
from the architecture library pool such as libx86_64arch (Step 2 in the
figure). With this in place, the user saves the configuration and types
“make” to build the unikernel (Step 3) and “xl create” to run it (Step 4).
A note on the ABI/API exposed to the application: because Unicore allows
for customization of the unikernels, the ABI (or API since there is no
kernel) would be custom, that is, defined by the libraries the user
selected. Having said that, it would be perfectly possible, for instance,
to build POSIX-compliant unikernels with it (e.g. similar to Rump, but in
principle with much more specialized OS layers).

Finally, it is worth pointing out that we use the term application
loosely: another clear target for Unicore is the building of
runtime-specific unikernels (e.g. a unikernel able to run Python or OCaml
scripts as is the case with MirageOS).

Relevance to Xen and its Community
Unikernels are important to a number of areas relevant to the Xen
community, including IoT, automotive, stub domains, and driver domain/dom0
disaggregation. Unicore could help boost the progress in all of these
areas by quickly providing the necessary tools to create  unikernels for
them. For instance, for a driver domain, the user would include the
“library” containing the relevant hardware driver and corresponding
back-end driver, and in principle Unicore would take care of the rest.

In addition, Unicore could eventually replace Mini-OS, providing a
cleaner, more stable and flexible base from which to build unikernels for
projects (the modularization of Mini-OS is in fact already taking place).

Current Status
Unicore is at an early stage. For now it includes some base libraries with
code extracted from Mini-OS as well as a build tool inspired by Linux's
KConfig system. Unicore is currently able to build "hello world"
unikernels for Xen and Linux user space on x86_64 and ARMv7.

The reason behind making Unicore a Xen sub-project project is to (1)
bring the existence of Unicore to the attention of the Xen community
and to outside world; (2) to attempt to harness interest and
potentially development cycles from people and companies interested in
unikernels; (3) to concentrate maintenance resources from people
interested in unikernels within the community; and (4) to have a legal
entity behind the project.

The main license of the run-time components of Unicore will be a 3-clause
BSD license, unless there is a good reason not to use it (e.g. we may
import 2-clause BSD licensed code from Mini-OS, which we would *not*
anticipate to change). The Makefile system would be licensed under GPL v2
or later as we want to be able to use KConfig functionality from

Required Infrastructure
The official repositories should be created on
[http://xenbits.xenproject.org/] under `unicore.git`. There should be a
main repository for the core unicore implementation and additional
repositories for some more advanced extension libraries (e.g., lwIP,

### Main repository


### Repositories for extension libraries

Repositories for additional libraries that are supported by the Unicore
project should exist under a separate directory:


For example:


### Mailing list

In the beginning we would use the MiniOS mailing list
(minios-devel@xxxxxxxxxxxxxxxxxxxx). When we get traction with Unicore we
could consider splitting that traffic onto a unicore mailing list.

Dr. Felipe Huici
Chief Researcher, Networked Systems and Data
Analytics Group
NEC Laboratories Europe, Network Research Division
Kurfuerstenanlage 36, D-69115 Heidelberg
Tel.     +49
(0)6221 4342-241
Fax:     +49
(0)6221 4342-155

NEC Europe Limited Registered Office: NEC House, 1
Victoria Road, London W3 6BL Registered in England 2832014

Attachment: PROPOSAL Unicore.pdf
Description: PROPOSAL Unicore.pdf

Attachment: unicore-oneslider[1].pdf
Description: unicore-oneslider[1].pdf

Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.