Assess the performance of MPC consensus approaches

The new YandaSoft LSQR calibration solver scales as the number of parameters squared and allows for extra constraints to be added between otherwise independent systems, such as enforcing smoothness in frequency, time or direction on the sky. Work is ongoing on a parallel version that allows different systems to run in parallel with a minimal number of coupling parameters distributed between solvers.

As this solver works well at scale and is a form of MPC consensus, its performance (scientific and computing) should be compared with similar algorithms that also employ forms of consensus between compute islands.

Hardware available for tests:

• Dell PowerEdge M630 cluster with dual 10-core Intel Xeon E5-2660 V3 nodes running at 2.6 GHz with 25 MB cache and 128 GB of memory. Node interconnect is FDR Infiniband
• Bracewell cluster with dual 14-core Intel Xeon E5-2690 V4 nodes. Each node has 256 GB of RAM, 1TB local SSD and 4 NVidia Tesla P100 GPUs. FDR10 InfiniBand interconnect.
• If useful: SGI UV3000 cluster with 10-core Intel Haswell nodes and 8TB of global shared memory (SGI NUMAlink interconnect).

We will also do tests for ASKAP at the Pawsey centre on the ASKAP Cray XC30 cluster with dual 10-core Intel Xeon E5-2690v2 ‘Ivy Bridge’ nodes running at 3.00 GHz with 64 GB of memory and Cray Aries node interconnect.

Other architectures may be available via cloud services.