[Beowulf] Win64 Clusters!!!!!!!!!!!!

Tue Apr 10 05:47:48 PDT 2007

On Mon, 2 Apr 2007, Ryan Waite wrote:

> Yeah, that marketing slogan wasn't too great since it sounds like we're
> solely taking HPC mainstream. Ouch. Of course, that isn't the case.
>
> Where would people on this list place the credit for HPC going
> mainstream? If I had to pick one source, which is unfair to do, it would
> be the rich array of applications that exist, from BLAST to Fluent, from
> Ganglia to LSF, from PERL to Matlab.

Sorry, I just spent the weekend getting snowed and sleeted on while
camping at the beach -- froze my **s off, but caught many blues out of
the surf and had a great time.  So this is bizzarely time-delayed
(doubly so, from the datestamp above).  But I couldn't resist.

In my opinion, the place where HPC went COTS/mainstream even before "the
beowulf" was when Vaidy Sunderam and Al Geist (in Jack Dongarra's group
IIRC) wrote PVM at Oak Ridge in 1989.  Before that people used clusters
of workstations as rsh-based or sneakernet job distribution systems (did
a bit of both myself, sneakernet back into the early-to-mid 80's with
original IBM PCs and our first 286 boxes, the latter with Sun 4/110's,
386i's, SGI Irises, on thinwire).  And I claim nothing special there --
anybody with big computing needs and multiple computers to work with has
been using "clustering" basically forever, all the way back to Enigma
days (which wouldn't be a bad place to begin in the non-COTs category).

With PVM, building a distributed parallel supercomputer made out of COTS
computers using an OTC network for IPCs was reduced to basically
selecting unixoid (tcp/ip/rsh supporting) hardware and cabling it
together, or (as in my case) "discovering" that an existing network of
workstations already installed so that researchers could have compute
resources and access to the rapidly growing list of resources available
via "the Internet" (capital "I") WAS de facto a parallel supercomputer
which for a variety of tasks could easily give you far, far better
performance, for "free", than $12M Cray vector machines where if you
begged nicely they'd give you "100 hours of Cray time" as if that would
make everything wonderful.

Frankly, PVM was the sine qua non of the original beowulf.  One could
easily build immensely large compute clusters out of Sun workstations --
I got up to roughly 130 machines all over Duke's campus working 24 hours
a day on a major compute project back in 1995 -- or pretty much any
other Unix boxes.  One could even use a heterogeneous network of
machines -- Vaidy demonstrated this with some lovely graphs when he gave
a talk at a workshop I attended back in 1992 (I think) where he showed
parallel speedup in a computation using a Cray, a pile of DECstations,
some Sun workstations, and a few other things I cannot remember (maybe
AIX boxen?) all at once.  WITH PVM, building the original beowulf was a
matter of adding network drivers to the linux kernel and working to
speed up the current somewhat pedestrian 10 Mbps ethernet channels
available for PCs until faster networks came along (as they always do).

Without PVM, there would have been no point, as only true masochists
would ever write a "real parallel program" on top of raw sockets.

Applications are not what made parallel supercomputers mainstream at
all.  Most researchers have always written their own code, or used a set
of code that was originally written by a single group and then
disseminated throughout a community as a community resource.  These
applications could not have been written without a base to write on,
and without PVM originally to "create" (or enable) a true COTS cluster
with message passing capabilities, there would have been no hardware
base to motivate the porting of MPI to open the door to moving all of
the existing MPI applications to COTS clusters (well, there would have
been, sure, but it would have been delayed by years instead of an idea
waiting to happen).

Anyway, the short history above is why the marketspeak implying that the
advent of Windows Clusters somehow signals clustering becomming
"mainstream" is so (amusingly) offensive to long time clustering people.
Clustering was "mainstream" back in the days of Von Neumann and Alan
Turing.  Gene Amdahl wrote a seminal paper on clustering back in 1967.
To quote from Greg Pfister's book (as quoted on wikipedia:-)

"Virtually every press release from DEC mentioning clusters says 'DEC,
who invented clusters...'. IBM didn't invent them either. >>Customers<<
invented clusters, as soon as they couldn't fit all their work on one
computer, or needed a backup. The date of the first is unknown, but it
would be surprising if it wasn't in the 1960s, or even late 1950s."

emphasis my own -- and note that it is clearly documented that
sneakernet as a clustering technique dates back at least to Enigma and
the mid-1940's.

Clustering has clearly been "mainstream" in the research world at least
since 1992-1993 (when PVM-evangelists spread the word that it was now
possible to write portable programs for TCP/IP based virtual
"supercomputers" on COTS hardware) and it has been IDENTIFIABLY
mainstream with its own "brand name" since the beowulf project created
the beowulf list and people who were co-inventing it on their own anyway
all joined together to create a powerful and consistent support network
to ease the way for newbies.  The top500 list over the last decade or so
(much as I generally deplore it) clearly marks out "a" visible history
of the process of clustering achieving dominance, and there is hardly a
mention of Microsoft on it anywhere.

     rgb

>
> Thanks for the feature requests, Joe.
>
> Ryan Waite
> Group Program Manager - HPC
> Microsoft Corporation
> ryanw at microsoft.com
>
> -----Original Message-----
> From: beowulf-bounces at beowulf.org [mailto:beowulf-bounces at beowulf.org]
> On Behalf Of Joe Landman
> Sent: Monday, April 02, 2007 12:07 PM
> To: Douglas Eadline
> Cc: Erik Paulson; beowulf at beowulf.org
> Subject: Re: [Beowulf] Win64 Clusters!!!!!!!!!!!!
>
> Douglas Eadline wrote:
>
>> I believe that if we do not protect against revisionist history, then
>
> [...]
>
> you mean like how now with WCCS2k+3 clustering and HPC is *now*
> (suddenly magically spontaneously) "mainstream" ?
>
> This is just something I personally take issue with.  The entire
> explosive growth of clustering has driven HPC hard into the mainstream.
>  This happened long before it was a glimmer in their eyes.  6+ years of
>
> explosive growth, going from noise in the statistics to dominating the
> statistics.  Then along they came with WCCS2k+3.
>
> Their entry is late into the cycle.  And if you listen to the comments
> of the senior execs, it makes one wonder how committed they are to HPC
> and clusters as compared to how committed they are to battling Linux.
>
> This is not to diminish their efforts.  WCCS2k+3 is likely reasonably
> good for some subset of groups.  Microsoft has some good people there,
> and playing with the W2k+3 x64 on our JackRabbit unit was fun.  They
> still need a real POSIX subsystem, and hopefully, someday, they will
> give in, and get cygwin or mingw to be fully supported/shipping using
> their compilers/tools.
>
> Though I expect to see airborn and stable flight from porcine critters
> about the same time.  Too bad, as that would likely ease
> adoption/porting issues.  Tremendously.
>
>

-- 
Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu