[Beowulf] How Can Microsoft's HPC Server Succeed?

[Greg Lindahl]

On Fri, Apr 18, 2008 at 04:22:13PM -0700, John Vert wrote:

> V1 ("Windows Compute Cluster Server 2003") uses
> WinsockDirect. High-speed interconnects like Infiniband plug into
> this stack through the existing WinsockDirect provider interface.

Ah. It's still considerably slower than real native support.

> V2 ("Windows HPC Server 2008", coming soon) introduces a new
> provider interface called NetworkDirect which maps much better to
> the hardware. So far we are seeing excellent performance and the 2
> microsecond latency quoted earlier is one example.

OK, so how about a test sensitive to overhead? Like mpi-multibw?
Latency isn't a very good test because most programs don't just
ping-pong between 2 processes. They send to a bunch of neighbors at
once.

I could easily believe that it's better than V1. But most people
want to know if it's as fast as what they're used to.

> Our V2 job scheduler also has a lot of performance improvements. If
> you care about how long the scheduler takes to submit, allocate,
> reserve, and activate a 1,000+ CPU job, I think you'll like
> that. This is really nothing to do with "Linux clusters" as it's
> largely a job scheduler issue and most job schedulers support
> multiple platforms.

Then why the slam against slow Linux cluster startup times in a blog
and presentations?

-- greg