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

Re: [MirageOS-devel] [Minios-devel] [Xen-API] [RFC] Unicore Subproject Proposal

On 15.09.2017 11:35, Anil Madhavapeddy wrote:

On 15 Sep 2017, at 09:36, Simon Kuenzer <simon.kuenzer@xxxxxxxxx <mailto:simon.kuenzer@xxxxxxxxx>> wrote:

Hey Anil,

On 13.09.2017 12:11, Anil Madhavapeddy wrote:
On 11 Sep 2017, at 13:08, Simon Kuenzer <simon.kuenzer@xxxxxxxxx <mailto:simon.kuenzer@xxxxxxxxx>> wrote:

Just my 2 cents:
1. Is this academic project, or it have specific goals and areas of application? Would be good to have some practical use-cases and well formulated list of problems (we all feel these by guts, but...), it aiming to solve. IMHO that will help to prioritize functionality and get usable result faster :)

It is kind of both, however we aim a strong focus on real world problems: IoT, Mobile Edge Computing (MEC), Automotive, Virtual Network Functions (VNFs), and others. We have played with many Unikernels (ClickOS, Mirage, Rump, OSv, and others) and tried to apply them in the several areas. While doing this, we noticed that each area benefits differently from the properties that Unikernels give - which is great (e.g., instant boot times for MEC, high performance for NFV, resource efficiency for IoT). However, building and maintaining new Unikernels (as we did with ClickOS, MiniCache, and Minipython) is currently painful. Because of different focuses on properties and ported/implemented applications, most Unikernel today are bound to their own OS layers (e.g., ClickOS uses a different Mini-OS than Mirage). Each application requires a different subset of OS layers but also enables different optimizations of them.

In order to solve this, we came up with the Unicore proposal. But I agree with your suggestion at this point: It helps for the project start to focus on some initial areas. For now, I hope this is driven by the first contributors, and I have personally IoT in mind. Since the project goal is so ambitious, we should keep the long-term goal in mind from the beginning.

Thanks very much for kicking off this initiative. Maintaining a forked MiniOS has been a multi-year source of a maintenance burden for MirageOS, and we would love to be more aligned with an upstream version and benefit from new features such as (e.g.) HVM booting.

We have the same burden with ClickOS and all the other unikernels we have built. Features like HVM booting or support for different hypervisors, are always something that users ask for. Since many Unikernel projects struggle with this, I would like to have the maintenance effort of a common base concentrated. But we also learned that each Unikernel has own requirements: This is why Unicore has to provide full configuration flexibility. Only then, all Unikernel projects could really benefit from it.

So, I think we should all focus on the Unicore base and make our individual projects successful with it ;-).

This sounds good. It's worth thinking through the explicit differences in goals from MiniOS, to address Samuel's points.

Yes, I agree. I am going to have a second round through the proposal.

It seems to me that MiniOS should remain the primary "support kernel" for Xen-related activities, for instance as a stub domain support kernel. Unicore on the other hand explicitly tries to parameterise across its configuration so that it can be library linked to language runtimes more easily.

This split may help simplify MiniOS by removing some of the pseudo libc code that may be unnecessary outside Xen support functions, and also let us package up language runtimes in Unicore more easily via simple library package management and cross compilation.

It might be handled in this way during the project start, however I see Unicore more than just a set of base libraries to support language runtimes. The main intention is to support applications natively written to it and to support existing applications you may want to port. For these cases you need more than a minimal base, you want to select a libc, a certain scheduler, and most likely a TCP/IP stack, as well as filesystem support. These libraries should work together and should also work with the base libraries. Additionally, the requirements to the libraries might differ among the projects. Our content cache node Unikernel (called MiniCache), for instance, uses newlibc and lwIP. On top we have developed a single-purpose filesystem and web server code. Together we could achieve impressive performance results by using a single CPU compared to traditional web server applications running on a traditional OS [1].

In sum, the aim with Unicore is to have fine-granularity libraries so that it's flexible enough to accommodate a wide range of unikernels. Because of this, I see Unicore as having potential to eventually replace MiniOS.




[1] http://dl.acm.org/citation.cfm?doid=3050748.3050757

Simon Kuenzer
シモン クゥンツァー
Research Scientist,
Networked Systems and Data Analytics Group
NEC Laboratories Europe, Network Research Division
Kurfuerstenanlage 36, D-69115 Heidelberg
Tel.     +49 (0)6221 4342-264
Fax:     +49 (0)6221 4342-5264
e-mail:  simon.kuenzer@xxxxxxxxx
NEC Europe Ltd | Registered Office: Athene, Odyssey
Business Park, West End Road, London, HA4 6QE, GB
Registered in England 2832014

MirageOS-devel mailing list



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