Question about custers
hahn at physics.mcmaster.ca
Fri Feb 7 16:28:34 PST 2003
> I wanted to ask if there's a way of theoreticaly calculating a cluster
> power by a mathematical formula, basing on the nodes procesor type, ram,
> etc.? Assuming also that the components of each node can be different.
no, there is not. at least not any way that's both intelligible and general.
intelligible but not general is easy: run a fixed application on all
the clusters to be compared.
general but not intelligible is also easy: run as many applications as
you can find on all clusters. for N applications, the result is close
to N-dimensional, and therefore difficult to comprehend for anything
except trivial (N <= 4) exercises.
hmm, that's actually an interesting take: instead of taking the
geo mean of spec scores, run them through PCA or ICA. this will
decompose them into new synthetic benchmarks ("eigen benchmarks")
which are orthogonal or independent. I guess it would also provide
a sane way to predict your non-spec app's performance on new
machines if you have some scores on spec-known machines...
anyway, PCA/ICA often provides dimensionality-reduction, which would
make it easier to blather about "cluster power".
on the other hand, I don't see any reason not to just sum the clocks
of each node for "cluster power". or use BTU's. or sum the bandwidth
of each node's cache and memory bandwidths. or the inverse of their
memory latencies. oh, hell, just sum each node's bogomips!
regards, mark hahn.
PS: I stood behind a wall of 512 1U Xeon/2.4's today, and was thinking
that it made a really great whole-body, parka, hat, boots dryer.
seems as good a definition of "cluster power" as any...
More information about the Beowulf