[Beowulf] Re: newbies' dilemma / firewire? (Hahn) (Will)
Ed Karns
edkarns at firewirestuff.com
Fri Mar 10 08:58:31 PST 2006
On Mar 9, 2006, at 7:32 PM, beowulf-request at beowulf.org wrote:
> Infiniband with DDR is already at 20Gbps over CX4 copper
" The 4X InfiniBand protocol extends the existing 1X protocol by
supporting up to four 2.5Gb/sec dual-simplex connections for an
effective duplex transmission speed of 10Gb/sec. ... " (From: http://
www.lecroy.com/tm/products/ProtocolAnalyzers/infiniband.asp?menuid=62 )
I did not mean to imply that this performance level was not possible
or good or valuable ... just not "cost effective" considering:
energy budget (line length v. energy consumed), upper limits,
etc. ... Infiniband (and others) can be pushed even further over
silver conductors (or carbon or super conductors) ... Heck, you could
make the above case for any protocol performance pushed to the upper
limits of meat space (physical realities). Consider a protocol as
being reliable or workable, when changing the means of connectivity
to another medium. Optics can open up an order of magnitude
performance gain = electrons v. photons, metal v. glass (or air or
other) ... beyond a bus speed of 100 or even 500 Gbits / second @ a
lower energy requirement ... Incremental performance improvements of
late over metal conductors merely prove the need ... metal conductor
data transmission is falling behind Moore's "law".
This is going to become very important, very soon as more advanced,
dramatically improved performance Beowulf systems are built. My vote
would be for the most hardware / firmware efficient protocol
considering energy budget v. performance v. the space allowed. (Cray,
IBM, et al, currently build clusters that have hundreds of horsepower
devoted to heat dissipation ... just because they all use metal
conductors for the bus. Imagine having a twin engined aircraft
running inside your server farm ... being something most of mortals
can not afford, let alone survive.)
I still pose the question: which hardware protocol would be optimum
for tight clusters of processors sharing a common bus (or other local
hardware network) in a Standard Temperature & Pressure environment?
"Raw Data" transmission ala legacy serial or parallel or SCSI or
other ... ?
"Packet Switching Data" transmission ala Ethernet or USB or FireWire
or other ...?
A "yet to be determined data" transmission methodology /
topography ... ?
Ed Karns
FireWireStuff.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20060310/f9ca7ca9/attachment.html>
More information about the Beowulf
mailing list