HEPiX Virtualisation Working Group Meeting

Monday 8 Feb 2010, 16:00 → 17:00 Europe/Zurich

Phone conference

Tony Cass

<sl>

Dial +41227676000 and enter access code 0186715 followed by #

Key ##1 to mute/unmute your phone </sl>

HEPiX Virtualisation Working Group Meeting Minutes Monday 2010-02-08

Present (in no particular order):
James Adams, RAL
Ewan MacMahon, Oxford
Keith Chadwick, FNAL
Martin Bly, RAL
Sonder Klous, NIKEF
Philippe Olivero,  IN2P3
Matthieu Puel, IN2P3
Ulrich Schwickerath, CERN
Ricardo Silva, CERN
Sebastien Goasguen
Thomas Finnern, DESY
Dave Kelsey, RAL
Michel Jouvin, GRIF/LAL
Andrea Chierici, INFN Bologna
Andrew Washbrook, Edinburgh
Alessandra Forti, Manchester
Gary Stiehr, U Washington/ Genome Centre
Owen Synge, DESY
Ian Collier, RAL


If there are any omissions or errors please let me know (Ian Gable Tony opened by highlighting people for particular work areas:

Dave Kelsey  and Keith Chadwick, Image Generation
Owen Synge, Distribution
Sebastien Goasguen, Contextualization
Andrea Chierici, Multiple Hypervisor


1. Image generation

Keith:

• There is a wide variety of topics to be discussed.
• There are a variety of things that the hypervisor can do to limit what the image can do.

Dave:

• Thinking as chair of Joint Security Policy Group(JSPG) we need to develop a trustworthy method for generation of virtual images.
• He would like to develop a written policy document to define just what a trustworthy system for the generation is.
• Document needs to be developed in consultation with site managers, security people, and VOs

Sonder Klous:

• Are we going to run trusted images only?
• Tony responded that there unless there was a process to create trustworthy images then the default would remain; the default is to run nothing unless it was generated at the site.

Owen:

• Owen proposed that there is some process by which we can generate an image (for example a list of RPMs which would be used).
• Tony commented that this was not what he had intended. It may be that each site will not have the capability to generate the needed image (missing RPMs or experiment software).

Timing of the work:

• Dave mentioned that there is a JSPG meeting in March. He asked if this in the timescale of the group.
• Tony commented that he would like to move on this much before the end of March.
• Dave asked how prescriptive  do we want to be in the policy? He commented that we should define what we want to be able to do before  defining the security policy.

2. Distribution

• Owen commented that a particular site must be able to perform a security investigation on an image and chose not to run at their site.
• Tony commented that if an image is unsuitable for one site then it's likely unsuitable for all sites.
• Owen agreed with the above.
• Tony commented that this is we need an image revocation mechanism, possibly something like a certificate revocation list.

• Owen commented that he would like to separate image distribution into two levels, 1) central repository to the site 2) site to the workernode.
• Tony commented that we should be agnostic about the second level.

• Owen asked if we want to transmit full images or difference images. Owen prefers the simple mechanism of full images.
• Martin commented that we don't need to be prescriptive with how the sites get images, as long as the image is verifiably identical.

• Ewan MacMahon commented that this assumes that the site administrator are pulling the images from central repository and not that users will be providing their own images.
• Tony pointed out that the assumption of the working group is that we will have some small number of providers of images, and those images will be pulled to sites. The cloud model here is something, from HEPiX experience, that sites are not ready for *today*, but maybe in the future. We need to focus of the first step, which is trusted images.

• Owen commented that there might be demand for users to download images to their own machines, thus creating more demand then we had anticipated.

• Owen asked about people general feelings about using BitTorrent as a mechanism for distributing images
• Ewan MacMahon commented that central networking at organizations like his often have a problem with BT. However he himself thinks it's a excellent protocol.
• Martin Bly commented that there might be trouble at RAL
• Tony commented that if the case can made for it's profession use then it should be possible to overcome the difficulties at CERN
• Owen commented that he believes that he would be successful in getting support for Bittorrent from the appropriate people at DESY.

 

3. Contextualization

Sebastian:

• The typical mechanism for contextualization is to attach a disk to the VM and then run some scripts which are contained in that disk.
• Both Nimbus and OpenNebula do contextualization. CERN at the moment uses  Open Nebula to do contextualization.
• The identity of the machines is not permanent and hence is not configured exactly like a permanent machine.
• Machines have have typical lifetimes of hours to days, sometimes, but rarely, have lifetimes of weeks.

 

Others:

• Owen commenting that we need to gather a list of things that we want to be contextualized.
• Tony suggests that rather defining what we want to contextualize we should define what we must not change in the image.
• Summary: We want to avoid making contextualization changes that defeat the purpose of having VMs. The classic example of a bad contextualization thrown around was replacing the version of python

• Ewan suggested rather then passing a executable script to the image we pass configuration parameters.
• Sebastian commented that he liked the idea but it is not possible to prevent site admins from modifying images, we must focus on what we want to achieve.

4. Multiple Hypervisors

Andrea Chierici:

• The idea is that we can seamlessly run the image on different hypervisors
• Do we need to consider more then just Xen and KVM?
• There was a comment (name unknown) that we should support all hypervisors.
• Michel suggested that we limit to 2 or 3 hypervisors purely for practical reasons.

Others:

• Ewen commented that we should focus on hypervisors that have embedded kernels.
• Owen commented that he considers the kernel part of the contextualization.

5. Next meeting

• Owen, Martin, Ewan MacMahon to discussion on distribution.
• Owen would like to have a discussion on the mailing list on which things should be contextualization
• Andrea will create doodle poll on hypervisors.
• Ewan suggested keeping discussions on the mailing list as much as possible even when the discussion is within a particular sub group.

16:00 → 16:05 Previous Minutes

Minutes http://indico.cern.ch/getFile.py/access?contribId=0&resId=minutes&materialId=minutes&confId=80869

16:05 → 16:55 Work area discussions

16:55 → 17:00 Agenda for next meeting