Re: Skywriting + mirage

[Anil Madhavapeddy <anil@xxxxxxxxxx>]

On 23 Feb 2012, at 09:33, Sebastian Probst Eide wrote:

> Dear miragers,
> 
> In the "Turning down the LAMP" paper, section 5, it is mentioned that mirage 
> is a good match as an execution platform for skywriting (and by extension 
> CIEL). Has there already been done any work in that area?

Hey Seb,

Yes, there's been ongoing work in that area since the paper. The problem with 
specialising CIEL tasks (or indeed, most data-intensive processing) is that the 
interconnects become increasingly inefficient as the worker kernels become more 
fine-grained.

Malte did some initial work on this to demonstrate the problem:
http://www.cl.cam.ac.uk/~ms705/paper-sfma.pdf

A bunch of us have been developing a more efficient 'elastic' I/O interconnect 
for connecting VMs:
http://anil.recoil.org/papers/2012-resolve-fable.pdf
http://anil.recoil.org/papers/drafts/2012-usenix-ipc-draft1.pdf (short paper 
version draft)
http://anil.recoil.org/talks/fosdem-io-2012.pdf

Meanwhile, David Scott has been rewriting the Xen toolstack to better support 
low-latency microVMs (aka 'stub domains').
http://github.com/djs55/xen-api (cooper branch, unreleased code, Dave's on this 
list).

We're integrating this into Mirage (with multi-language support) at the moment. 
I'll have an update on that in a week or two, as it relates to the upcoming 1.0 
release plan for Mirage.

-Anil