[nSLUG] Atom vs Xeon

Stephen Gregory nslug at kernelpanic.ca
Fri Apr 30 21:27:47 ADT 2010


Daniel Morrison wrote:
>
> http://www.technologyreview.com/computing/22504/
> 
> FAWN - Fast Array of Wimpy Nodes

This idea comes up constantly but products rarely get out the door, or
see widespread use. The researchers pushing these systems typically
underestimate the real world working load, over state bottlenecks,
forget to include the power and system cost of the whole system, and
make the wild assumption that their little system can be scaled up. It
is interesting what these researchers are trying to achieve, but it is
not working in the real world.

The 21 node array in the article can serve 31k queries per second (364
queries/watt * 85 watt). A Sun UltraSparc T2 based system can achieve an
order of magnitude more queries per second. A T2 is a properly designed
FAWN on a single chip. A T2 system running memcached achieved 250k to
530k queires per second depending on object size. As the Geode base FAWN
was designed to compete with memcached I am guessing the comparison is
valid. A Netra T5220 came with a 600w power supply. Assuming worst case
power usage that is still 400 to 880 queries/watt. It is likely much
better then that as the T5220 almost certainly uses less then 600w.

http://blogs.sun.com/zoran/entry/scaling_memcached_500_000_ops


> What does "fixed power requirements of the data center" have to do
> with "throughput"? Throughput of what? Electrons?

The entire article is about throughput (of search queries) versus power.
 As these servers don't stand alone the power requirements of the entire
system (e.g. data center) is relevant. The power requirements of the
data center are part of the reality that researchers pushing FAWN type
systems tend to ignore.


-- 
sg




More information about the nSLUG mailing list