incubator-general mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "David O'Hallaron" <>
Subject Re: [PROPOSAL] Tashi
Date Tue, 15 Jul 2008 13:55:40 GMT

>  * The sponsoring entity is the Incubator so I'm guessing you're shooting
> for graduating as a TLP. What kind of interactions do you foresee with
> Hadoop for example?

We talked with Doug Cutting about whether to shoot for a TLP or a
subproject under Hadoop. We decided ultimately to go the TLP route
because Tashi and Hadoop are at different levels in the stack.
However, we see Hadoop as one of the important applications running on
Tashi virtual clusters, so the two projects are quite complementary.

>  * IIC Tashi is only about management, not about the underlying
> storage/computing technology. Is that correct? And if so which ones do you
> plan to integrate with?

Yes, that's correct. We plan to integrate with the major VMMs, such as
Xen, Linux KVM and VMWare. For storage, we're looking at integrating
with HDFS, pVFS and later pNFS when it becomes more mature.An
important goal is to provide the hooks and interfaces that allow any
DFS and VMM vendor to integrate with the system.

>  * I can't help asking for more technical details. What's the
> implementation language for your POC code? What are the non-proprietary
> interfaces you're thinking of?

The POC code is a couple of thousand lines of original Python code.
Major components are cluster manager (cm), which runs on one of the
cluster nodes, a node manager (nm), which runs on each of the other
physical cluster nodes, a simple db on the cluster manager for
configuration data, and some client utilities.

We're really thinking hard about interfaces now, but don't have clear
definitions yet. The kinds of things wer are  thinking about are
interfaces between the client and cm and cm and nm for manipulating
starting, starting, and migrating vms, interfaces for some kind of
event/messaging system for monitoring and reporting system state to
the cm and client, interfaces between the cm/nm and storage system to
allow the cm to do storage-aware scheduling of vms, interfaces to
power management and system management hardware features, and possibly
interfaces for federating different Tashi clusters.



-- David O'Hallaron
-- Director, Intel Research Pittsburgh
-- Assoc Prof of CS and ECE, Carnegie Mellon University

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message