Intel® Math Kernel Library 11.3 Update 4 Developer Guide
While it is relatively easy to get high performance of the HPL test on a single node, it is more complicated in a cluster. To achieve high performance of the test in a cluster, follow these steps, provided all the needed installations are done on each node:
Reboot all nodes.
Ensure all nodes are in identical conditions and no zombie processes are left running from prior HPL runs.
To do this, run single-node Stream and HPL LINPACK or the Intel MKL program nodeperf on every node. For more details, see also Optimizing the Result on a Cluster. Ensure results are within 2% of each other (problem size must be large enough depending on memory size and CPU speed). Investigate nodes with low performance for hardware/software problems.
Check that your cluster interconnects are working. Run a test over the complete cluster using an MPI test for bandwidth and latency. The IMB and Microway* Linkchecker tests of Intel MPI are two of many benchmarking methods that can be useful for this step.
Run an HPL benchmark on pairs of two or four nodes and ensure results are within 4% of each other. The problem size must be large enough depending on the memory size and CPU speed (for example, refer here).
Run a small HPL workload over the complete cluster to ensure correctness.
Increase the problem size and run the real test load. Rerun at the real size at least three times.
In case of problems go back to step 2.
Each step is important, so skipping a step like 4, you can waste a lot of time.
Before making a heterogeneous inter-node run, always run its homogeneous equivalent first. If you are using Intel Xeon Phi coprocessors in an offload mode for intra-node heterogeneity, first run on the Intel Xeon processors alone.
|
Intel's compilers may or may not optimize to the same degree for non-Intel microprocessors for optimizations that are not unique to Intel microprocessors. These optimizations include SSE2, SSE3, and SSSE3 instruction sets and other optimizations. Intel does not guarantee the availability, functionality, or effectiveness of any optimization on microprocessors not manufactured by Intel. Microprocessor-dependent optimizations in this product are intended for use with Intel microprocessors. Certain optimizations not specific to Intel microarchitecture are reserved for Intel microprocessors. Please refer to the applicable product User and Reference Guides for more information regarding the specific instruction sets covered by this notice. Notice revision #20110804 |