| Monday, September 11 |
| 2:00 pm - 3:30 pm | |
| |
Grids and System Virtualization
(90 mins)
Franco Travostino, Yaron Haviv
(Grids and System Virtualization) Charter-Discussion BOF
View Participants
System architectures have evolved along two complementary dimensions:
(i) "virtual machine" techniques (e.g., xen, vmware, etc.) which allow fully featured operating environments (e.g., Linux, Windows) to coexist within the same processor and physical memory, unbeknownst to one another. From an application standpoint, such environments are undistinguishable from native environments. Virtual machines enable dynamic isolation, consolidation, and migration capabilities, resulting in lower TCO and efficiencies in capital/operational costs.
(ii) virtualized I/O resources that represent slices of the overall I/O estate of the physical system. Such resources include: network endpoints, sessions in a TCP Offload Engine (TOE), queue pairs and virtual lane associations in an Infiniband Host Channel Adaptor (HCA), IPsec Security Associations (SA), etc. The demand for determinism, low latency, high throughput, and greater security prompts for increasing sophistication of these virtualized I/O resources.
We refer to the combination of (i) and (ii) as virtual systems. Virtual systems come with a set of tools to set up, tear down, re-provision, and migrate a virtual machine and its budget of I/O resources. In short, there are tools that assist in all aspects of the Virtual System Lifecycle (VSL).
Within a Grid, it is quite possible that VSL might be entirely assimilated within traditional system management. In other words, instantiating a virtual machine and populating its I/O resources is no different than powering up a physical board and provisioning its network interfaces. On the other hand, there are cases where job description, resource matchmaking, job scheduling, replica management, etc. thrive on directly driving one or more aspects of the VSL. For instance:
a. I require that my workload be carried exclusively by just-in-time instantiated virtual systems (e.g., for greater isolation, avoidance of malware, etc);
b. I want my workload to be executed by virtual systems because I anticipate the need to perform suspend/resume operations on the workload prior to its completion;
c. I want to migrate my running workload, inside or outside a data center, because I need to optimize data/instrument locality, harvest "track the moon" efficiencies, or increase geographical spread to meet regulatory requirements for business continuance, disaster recovery.
In these cases, the Grid infrastructure must be cognizant of virtual machines capabilities and virtual I/O capabilities, yet without having to know about their actual implementation. As well, the Grid infrastructure must be prepared to operate them in unison, to avoid gaps in performance or in the user experience.
With this BOF, we propose to launch activities that:
(i) within existing Grid standards, verify that requests for systems with certain capabilities might be satisfied by either physical systems, or by the infrastructure creating and managing virtual systems;
(ii) analyze use cases similar to a), b), c) wherein the Grid infrastructure is seen participating in the VSL (as opposed to the VSL being part of traditional system management in a data center);
(iii) define the Web Services interfaces that allow a Grid infrastructure to access selected aspects of the VSL in an implementation neutral fashion.
We anticipate that the proposed activities will have a liaison to the DMTF's System Virtualization, Partitioning, and Clustering (SVPC) working group and the CDDLM group.
Location: 156
|
| |
| |
|
| |
| | Slides (PDF): Cees de Laat's presentation (power-aware computing) |
| | Document (PDF): DMTF presentation by Rob Goering |
| | Document (PDF): Introduction |
| | Document (Text): Meeting minutes (Cees de Laat) |
| | Document (PDF): Yaron Haviv's presentation |