Results

Current state

No measurements have been recorded. The performance statements elsewhere in this reference are predictions from the cost models in Computational Complexity and Machine Model, and are labelled as such where they appear.

This page holds the schema and the procedure, so that a run can be added without deciding how to report it.

Machines

Identifier	CPU	Cores	Peak (GFLOP/s)	Triad (GB/s)	Ridge point
(none recorded)	n/a	n/a	n/a	n/a	n/a

Each identifier expands to a full characterisation in the template given in Machine Model. A machine without a complete characterisation cannot host a result.

Kernel results

The schema for one kernel on one machine:

Field	Content
Kernel	Name and source revision
Machine	Identifier from the table above
Compiler	Vendor, version, complete flags
Sizes	The swept range, with the cache boundaries marked
Threads	Count and affinity
Statistic	Minimum or median, with dispersion
Achieved	Rate at each size
Bound	Roofline bound at each size
Fraction	Achieved divided by bound
Notes	Anomalies, conflict-miss signatures, frequency behaviour

Inner product

Machine	$n$	Threads	Achieved (GB/s)	Bound (GB/s)	Fraction
(none recorded)	n/a	n/a	n/a	n/a	n/a

Expected structure, from the model: a compute-bound plateau while both vectors are L1-resident, a step down at each cache boundary, and a memory-bound plateau at $I \cdot B = 0.125 B$ beyond the last level. See Inner Product.

Matrix multiplication

Machine	$n$	Variant	Threads	Achieved (GFLOP/s)	Fraction of peak
(none recorded)	n/a	n/a	n/a	n/a	n/a

Expected structure: the naive and interchanged variants flat and low, the blocked variant rising with $n$ toward a plateau near peak, and sharp dips at power-of-two $n$ unless the leading dimension is padded. See Matrix Multiplication.

Sparse matrix–vector product

Machine	Matrix	nnz	Threads	Achieved (GB/s)	Fraction of triad
(none recorded)	n/a	n/a	n/a	n/a	n/a

Expected structure: a fraction of stream bandwidth determined by the locality of the column indices, improving with a bandwidth-reducing ordering. Reporting GFLOP/s against peak for this kernel is misleading; the fraction of achievable bandwidth is the correct metric. See Sparse Computation.

Conjugate gradient

Machine	Problem	$n$	Preconditioner	Iterations	Time (s)	Time per iteration
(none recorded)	n/a	n/a	n/a	n/a	n/a	n/a

Both the iteration count and the time per iteration are required: a preconditioner is evaluated on their product. See Conjugate Gradient.

Adding a result

1. Characterise the machine using the probes in Machine Model, and add its row above.
2. Run the kernel under the protocol in Method: warm-up, individual repetitions, the stated statistic, and a second process to confirm.
3. Compute the roofline bound from the kernel's cost model and the machine's parameters.
4. Record achieved, bound and fraction across the swept sizes, not at one point.
5. Note anything anomalous: frequency behaviour, conflict-miss signatures at power-of-two sizes, scaling that saturates below the bandwidth limit.
6. If a figure could not be reproduced within its stated dispersion across processes, record it as unreproducible rather than averaging it into significance.

Why this page is empty

Filling it with plausible figures would cost nothing and would be worse than useless: the numbers would be quoted, and there would be no machine behind them. The reference states model predictions as predictions and measurements as measurements, and until a run exists there are only the former.

Related entries

Method, Machine Model, Examples, Conventions.