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

[Jim Lux]

Quoting Mark Hahn <hahn at mcmaster.ca>, on Sun 06 Apr 2008 01:58:01 PM PDT:

>> damage?  Presumably you have decent file system protection so that   
>> user A can't do bad things (or even see) user B's files.  All that   
>> happens is bad guy User A zaps their own stuff.
>
> this assumes that the windows admin knows enough not to ever run
> anything untrusted, or perhaps he's set up some fileservers/trees as
> readonly,
> or that users are not permitted to load their own executables, etc.
> I suspect that these things, which would be natural to any *nix admin,
> are not exactly second-nature to windows admins.  (disclaimer: I don't
> know any serious windows admins.)

Any serious windows admin would know all this stuff.  MS has had  
decent user access controls, etc., since WinNT came out. No enterprise  
scale system could work without it. Don't confuse the "admin-less"  
consumer model with the "admin-full" business/enterprise model.


The folks who get into trouble with Windows are the ones trying to do  
"admin-free" for a business, with the consumer strategies, when they  
shouldn't be.
  (I used to make a living untangling such things..)




>
>> Sure.. you let your cluster issue outbound network traffic to the   
>> big wide internet?  This is probably harder to actually allow than   
>> to prevent.
>
> huh?  I'm guessing the natural windows cluster organization is to put all
> the compute nodes on the full corporate/campus network.  it's not as if
> the windows world is really used to separating and tunneling GUIs over
> networks,
> at least not with the same level of naturalness as X and the usual   
> SSH tunnel.

I would say that the institutional windows world is as savvy about  
tunnels, remote consoles, partitioning networks as the institutional  
*nix world. If you look at MS's literature for CCS, they show the  
compute nodes on a private network, just as the typical *nix cluster  
is arranged.

And, for that matter, institutional Windows networks, on the whole,  
are probably managed a lot more tightly than institutional *nix  
networks.  Enterprise scale  windows (e.g. with domain controllers,  
SMS, etc.) gives pretty fine grained control over the user's  
workstation.   I'd venture to say that you'd have a lot more trouble  
dropping a "rogue" windows machine (i.e. if you ordered a computer on  
your own) into a enterprise network than a *nix machine.  You'd hit  
too many little hiccups (i.e. your login wouldn't authenticate against  
the domain, you wouldn't have access to shared resources, etc.)


This leaves aside things like SOHO (small office, home office) scale  
installations done by the "geek squad" or equivalent from a consumer  
electronics discounter.  Those are usually done as wide open share  
everything for everyone consumer installations (because it means they  
won't get a service call because "I can't see my shared drive on my  
husband's laptop")


>
>> Most clusters have a "totally inside the cluster" network that's   
>> only implicitly bridged to the outside world through the headnode.
>
> most _*nix_ clusters, yes.  but the whole discussion is windows, where
> users will naturally expect their job to see the same environment as
> their desktop,
> same filesystems, same graphics, same network access.

I don't know why a computational job would have graphics?  They run  
headless nodes in Windows just like *nix.  The headnode would have all  
the issues you address, but that's not the one one which the  
computation is being done, so all the penalty from running AV, etc.  
isn't as big an issue.


>
>> up the system.  But, also, recall the general model we were   
>> discussing.. smallish cluster to support some commercial   
>> application (say, a computationally intensive FEM code).
>
> one interesting fact is that the growth in cores of single machines is
> actually working _against_ the need for windows clusters.  who wouldn't
> rather just run jobs on a 32-core server, rather than screwing around
> with mpi on a cluster?
>
>> In this scenario, the cluster is basically sort of a "network   
>> attached appliance".  There are lots of network attached storage   
>> devices out there (e.g. from Maxtor) using some form of Windows as   
>> the OS.  They tend not to have AV stuff, just because the software   
>> on the appliance is fairly tightly configuration managed (i.e.   
>> nobody goes out running random programs on the NAS box).  It's just  
>>  not a huge threat.
>
> so you really think people will buy a packaged windows compute cluster
> preloaded with exactly one CFD code, and never be tempted to install
> other apps on it?  I think that's absurd, at least based on the kinds
> of things I see people doing with clusters.  the tool of the moment is
> constantly
> changing, even if the group isn't actually developing their own tools.

I do actually conceive of this happening.  We have full time  
mechanical engineers at JPL who probably use only one or two  
applications all the time (e.g. NASTRAN or other FEM codes). Likewise  
with some electromagnetic modeling tools.. an engineer using, say,  
Ansoft's products isn't likely to change to other codes, just because  
a)the existing product does what they need, and b)there's a HUGE  
learning curve to changing products in terms of building models and  
interpreting results.

Sure, they might want to install another code, someday, but odds are,  
they'd be buying another "cluster aware" shrinkwrapped application,  
which would be windows CCS compatible, and would fit within the MS  
software management scheme.

The target market for such clusters is NOT the researcher developing  
the codes or tinkering with multiple codes.  It's someone who needs  
more computational crunch than their desktop can give them, by an  
order of magnitude or two.  (i.e. rather than run 1000 segments in the  
model, you want to run 10,000, and it's a order (N^2) sort of problem)  
  The target market is someone who does not care what's inside the box.

  Jim