Martin Sivák

This talk should explain the internal mechanisms and differences between three schedulers that are currently used to plan workload placements in today’s datacenters and clusters.

The scheduling part of cluster management is commonly perceived to be complicated and close to dark magic. My goal is to describe the really simple concepts all three schedulers are based upon. I will then focus on the significant differences and strong and weak points of each of them in hope that I will uncover features worth of sharing between schedulers.

This talks should shed some light on how cloud and cluster scheduling works and facilitate the idea exchange in the scheduling area; both in the developer community and between developers and users.

The workloads and scenarios for virtual machines grow more complex every year. So do the interactions, availability, and performance requirements. All that requires the administrators to carefully plan where to start the VMs that depend on each other and/or specific hosts.

This talk will present the concepts that allow the administrator to express the rules for affinity between virtual machines and between virtual machines and hosts to form complex relationships that will cover for example:

• licensing rules - limiting group of VMs to only use certain hosts
• overhead - web + database VMs running together
• performance - eg. storage VMs running on hosts with better IO performance
• failover recovery - VMs returning to “their” hosts
• reservation - There is place for only one of us!

oVirt is an open source project for managing virtual data centers that will now help the administrator with exactly the above tasks. We have introduced the virtual machine affinity feature in the past and a huge improvement in that area is coming right now.

And the best part is that all this works in a fully dynamic environment with automatic conflict resolution and no manual management of host pinning rules, saving the administrator his precious time.

We were asked to come up with a solution to setting up a cluster where storage, management and VMs are running together on small number of nodes. This solution is currently known as hyper converged architecture. This kind of setup seems to be getting quite common in small deployments, but it was not possible to use it and maintain all oVirt reliability features until now. This presentation will be about the system design and issues we encountered while installing the minimal possible reliable cluster using three hyper converged hosts with oVirt as the software stack and Gluster for the storage - all running together in highly available setup across all three nodes.

The talk will cover the following topics:

• Identifying the failure points of virtual datacenter with separate management, host and storage nodes
• Highly available storage theory
• Highly available management theory
• Applying the high-availability in the context of an oVirt cluster
• Installation and management of hyper-converged setup
• Possible issues and things to solve