What should replace Linpack for ranking supercomputers?

The Linpack Fortran benchmark, which has been used to gauge the relative performance of workstations and supercomputers for many decades, is looking a little long in the tooth. So, some of the people who love Linpack and know it best are proposing a new benchmark - with the mind-numbing name of High Performance Conjugate …

[Michael H.F. Wilkinson]

An overhaul/replacement of LinPack is long overdue

Even the conjugate gradient benchmark, while very useful, does not say it all. Conjugate gradient problems are present in many optimization methods, but not in combinatorial optimization. Maybe an additional benchmark is needed for that. Some image and signal processing problems are not easily cast into an optimization framework at all, for these we would love to have a better benchmark.

[zooooooom]

Re: An overhaul/replacement of LinPack is long overdue

http://www.nas.nasa.gov/publications/npb.html

[Destroy All Monsters]

Will there be a book with all the TPM posts?

Suggested title "High-Performance Computing Memoirs found in a Bathtub: Missives from Before the Exascale Era"

Here is a fix btw:

The original Linpack from the dawn of time solved a desne matrix of 100 x 100 linear equations

That should be a dense matrix of 100 x 100 positions, for 100 linear equations to determine 100 unknowns.

Now, how about running some Lattice Gauge Theory simulations on them machines? An idea which goes back to Kenneth Wilson, who shuffled (shuttled?) off this mortal coil this very month. AFAICR problems of 17⁴ lattice points beach the mightiest machines dry. Nature apparently runs this thing at every point of the continuum, which I find extraordinary.

[GGorman]

What is the cost of a benchmark

The CG benchmark is a sensible choice as it tests the key general requirements of "high bandwidth and low latency, and tend to access data using irregular patterns". These are machine characteristics which have wide importance --- not just CG. Another interesting point in Jack's talk at ISC was the total cost of performing a benchmark. I forget the exact figure (left as an exercise to the reader) but it was in the order of $1m for the top machines! This tends to focus the mind on performing a minimum set of benchmarks rather than everything and anything.