|
|
Search the Sherrill Group
|
|
|
October 28, 2015: Hexagon has been upgraded to 15 nodes.
September 30, 2014: The Sherrill group has created the Hexagon
cluster for dedicated group computations. Each node is powered by an Intel
Core i7 hexa-core CPU, of the x930K series. Large-memory computations
are facilitated by the availability of 64GB per node. These resources
are in addition to similarly-configured workstations throughout the lab,
and access to the resources of CCMST and PACE. The cluster was constructed by Trent Parker, Matt Kennedy, and David Sherrill.
|
Cluster Specifications
Compute Nodes |
15 |
Cores |
90 |
Memory |
960 GB |
Scratch Space |
135 TB |
6-core 5930K nodes, 64GB @3.5-3.7GHz |
8 |
6-core 4930K nodes 64GB @3.6-3.9GHz |
6 |
6-core 3930K nodes 64GB @3.5-3.8GHz |
1 |
4-core 4790K head node 32GB @4.0-4.4GHz |
1 |
NVIDIA K40 GPU, 12 GB DDR5 RAM |
1 (hex9) |
|
Speedups due to Haswell FMA instructions
Two of the nodes feature new 5930K Haswell CPU's, which have a fused
multiply-add (FMA) instruction, allowing a potential doubling of
speed for matrix multiplies. In practice we see about a 1.8X speedup
in DGEMM using OpenBLAS, and about a 1.4X speedup in one of our timings
tests for CCSD(T) using the DF-CCSD(T) code in
Psi4.
Comparison of Haswell OpenBLAS (FMA-enabled, red), an older
(non-FMA-enabled) MKL library (green), and Haswell running OpenBLAS
compiled for Sandybridge (i.e., without FMA support, blue)
Sandwich Benzene Dimer DF-CCSD(T)/cc-pVDZ Wall Times
(seconds) Using 6 Cores
4930K (3.9GHz) MKL 10.3 |
1712 |
4930K (3.9GHz) MKL 11.2 |
1601 |
5930K (3.7GHz) MKL 10.3 |
1446 |
5930K (3.7GHz) MKL 11.2 (with FMA) |
1097 |
|
|
© The Sherrill Group
Georgia Institute of Technology