HT Re: [sldev] SL on Xeon processors
Dzonatas
dzonatas at dzonux.net
Fri Jun 8 23:00:18 PDT 2007
Peter Phillips wrote:
> Dzonatas,
>
> Are you alluding to the use of hyper-threading?
http://news.zdnet.co.uk/hardware/0,1000000091,39286539,00.htm
The reason why current OS thread schedulers fail on HT systems are (1)
dumb cache-awareness and (2) paired threads that need to use the exact
same resources that were not designed to be shared.
The methodology naturally follows more advanced software development skills.
http://en.wikipedia.org/wiki/Pair_programming
That wikipedia article starts to clearly state the idea, but it
immediately went into some other issue about political discipline.
The architects of the BlueGene/L have loudly stated that two physically
connected logical units are paired to do exactly two separate tasks. For
example, one accesses the network bus while the other number crunches.
In the BlueGene/L, there is 65,536 cores (two logical units per each
CPU). The machine does about 280 teraflops.
http://www.llnl.gov/asc/computing_resources/bluegenel/
This is by no means meant to compare people who sit near each other in
cubical style as an enabled HT multi-core processor.
Instead, Intel's 80x86 Terascale CPU has 80 of the x86 based processors.
It is rated around 1.8 teraflops for a single CPU. That is 80 logical
units per CPU.
If one replaced all the BlueGene/L CPUs with 80x86 Terascale CPUs, that
would be about 117,964 teraflops. We can expect something like that
performance increase in the next major supercomputer.
I allude the size of the 80x86 Terascale CPU to roughly the size of the
CPU that is currently in a typical workstation. It uses about the same
amount of power. The Xeon and others... similar size.
Intel's Nahelem, being partially derived from the terascale technology,
is only the "tock" of the particular microarchitecture series.
For SL, the "tick" after would be the realization of a filled in void.
=)
--
More information about the SLDev
mailing list