Quelques details en anglais sur les nouveautés de HPC Server 2008 (BETA 1)

HPC Server 2008 contains some noteworthy innovations:

1. It runs on  Windows Server 2008 only. There is no upgrade process. Wipe and replace.

2. It uses Windows Deployment Service (WDS), not RIS, which makes that "wipe and replace" much less painful :-) Multicast deployment is supported :-) although not the default choice.

3. The administration console offers a one-stop shop for deployment, administration, diagnostics and reporting. I know that many of you will be thinking, "just like Ganglia!". Well, that's the general idea. It is not full system center, but it is a very functional and efficient way to manage a HPC cluster. There will be a management pack for System Center 2007 as well.

4. It will offer head node fail-over from beta 2 onwards, thus eliminating a worrying single point of failure. This feature uses Server 2008 fail-over clustering, so it requires enterprise edition or better for the head node.

5. As a consequence of (4), we will support installing the head node on a sql 2005 cluster. In fact, we include sql 2005 express with the product but also support installations on pre-existing sql 2005 servers. You need not install the head node services on the sql machine either.

6. We have devised a new networking api to run along winsock direct, called Network Direct. the idea is to enable verb-based interaction with low-latency networking hardware, thus shaving off another couple of microseconds of latency, much like it happens with MVAPICH. In this release the only consumer of network direct is MSMPI. We're working with OEMs to write network direct providers / drivers. In beta 1, the network direct bits are present, but we have no released drivers yet. We are working with the usual suspects: Mellanox, Voltaire, Cisco and the Openfabrics consortium.

7. Powershell scripting is used for administration of common operations, along with the old v1 commands. In fact, those still work perfectly because we have maintained 100% compatibility with v1 COM API. Of course the v2 API exposes new functionality, but that deserves a post in itself. The powershell add-in is not complete yet. Job management commandlets work, but cluster management does not. Also, the old “job” command works, but does not expose the new features of the scheduler. There is a job2.exe in the bin directory that you can use instead.

8. The scheduler has been significantly enhanced for scalability and optimization. It deserves a post in itself, but here are some significant changes:

- ability to dynamically grow and shrink the pool of resources allocated to running jobs

- enforcing constraints on the basis of job templates, not just filters

- use of different units of allocation: core, CPU slot, node, depending on what your application needs

- biasing allocation algorithm towards memory or CPUs

- from beta 2 onwards, pre-emption of running tasks

9. Last but not least by any means, we are working with partners to support clustered file systems. CXFS, Melio and StorNext FS come to mind, being available on Windows now (2008 support is in the works).  More are coming: gpfs is definitely high on the priority list. Alas pvfs2 is not and we have no plan to produce a clustered file system of our own or port one and make it freely available.