LWN @ OLS: Alan's Linux-HA notes

Alan Robertson sent the following notes from the Ottawa Linux Symposium Clustering Working Group presentation. They are heavily based on the outline from Lars Marowsky-Bree, who chaired the presentation.

Goals of Clustering Framework:

To create:
    good HA products
    good HPC products
    viable/strong OSS project

Many different solutions spaces covered
Create opportunities for market niches
Encourage creativity and innovation
Maximize flexibility
Maximize reuse - esp. in future
Maximize ease of use/installation

Agree On further steps (how to proceed)

Identify components
Design APIs

externally-visible Standards (Application Programming Interfaces)
internal standards (Component Programming Interfaces)

Agree on above
Figure out how to rework current s/w into components. What can be reused, reimplemented (modified), written from scratch
Implement above
World Domination

How to proceed

Below is information from flip charts created during the discussion (3 hours wasn't really enough). Peter was busy taking notes during the discussion. I assume he'll have lots more information than I have here...

API conformance
CPI conformance

There was some discussion about how much distinction needed to be made between these two kinds of programming interfaces - and whether some APIs could also be CPIs. It was generally agreed that they could both be used by components. There was discussion about kernel-level versus user-level versions of these interfaces can both be made, and whether any particular component has to be implemented in the kernel or in user space.

Next on the flip charts was a picture I drew illustrating how components might interact, and the fact that there were multiple versions of various components. In the picture I noted that some proprietary systems might make calls to the APIs and/or CPIs to use the components. Also discussed the fact that proprietary systems probably won't fit the component model, but will certainly could easily have API/CPI conforming interfaces if they so desired.

Next flip chart...

IP Issues

clear
up front - every participant explicitly lays out relevant patent/IP issues before items are recommended.
agreements before we start
core vs non-core [AC]PIs and components

Goals:

IP is not a barrier for the creation of any core component.

Create a Policy Document with the following information up front

Expectations about clarity concerning IP issues from participants
Documentation license
Give the Doc/Spec a name, and trademark the name
OCF (Open Content something-or-other?)
No charge to write a compliant implementation (spec is free)
Possibly charge for certification (future)
Who controls the spec? -- Who's really in charge
Outline the general approach or philosophy to handling disagreements

Externally Visible APIs:

Much of this is brainstorming - not definitive...

comm

Ordered messages
barriers
etc.

System capability query (what components/interfaces are available)

This will be available at run time, and some systems may want it at configure time

Cluster membership
Group services

membership
version management (!)

Naming of "things"
Definition of a cluster (not really an API) -- not a peer-to-peer network
Authentication/authorization
DLM
Resource implementations
Quorum
Event services

ordered
unordered

NOTE: Check OSCAR for some of the configuration issues

Internally visible APIs

time? (see Eddie)
low-level cluster Remote Procedure Calls (RPC)

single machine/multicast
with replies (return value), no reply (return value)

low-level-communications
heartbeat
shared memory
marshalling/demarshalling (serialization)

must be unusually sophisticated - send arbitrarily complex lists/arrays, name/value pairs
self-describing data is mandatory

Self-configuration API for plugins...
fencing - reset mechanisms

high-level
low-level

Cluster Manager
Resource Manager
Logging
Initialization
Configuration repository
User Interface (configuration)

Some general comments on the reaction:

No one objected to the effort - and it appeared that there was a significant genuine interest.

Back to OLS 2001 coverage