Asking yourself how can you improve performance of your workgroup HPC workload? The answer is simple – move your workload to the IBM System x3750 M4! In this blog, I explain why.
The IBM System x3750 M4 server, built on Intel Xeon processor E5-4600 product family, delivers extensive set of HPC features to provide an optimal balance between performance, energy consumption, occupied space, and acquisition and management costs for a broad range of HPC workloads being either initially deployed or moved/consolidated from the existing implementations.
IBM Performance Lab conducted a series of benchmarks on IBM System x3750 M4 for different HPC workloads. The purpose was to analyze the performance scaling capabilities of the x3750 M4 itself under various HPC workloads, as well as to compare x3750 M4 with other dual- and quad-socket systems.
The performance guys ran the following benchmarks:
- Weather (WRF)
- Quantum Chromodynamics (MILC)
- Next Generation Sequencing (NGS)
- Computational Chemistry (GAUSSIAN)
- Computer Aided Engineering (FLUENT)
We explained the results of all these benchmarks in a new IBM Redpaper
. However, I want to share some observations for the LINPACK and WRF benchmarks with you here.
As we can see, the floating point performance of the 4-socket x3750 M4 with Intel Xeon E5-4650 processors is almost twice the performance of two 2-socket dx360 M4 servers with Intel Xeon E5-2680 CPUs.
Comparing to previous generation, single IBM System x3750 M4 can consolidate HPC workload from up to five 2-socket servers (we used dx360 M3 with Intel Xeon X5670 processors).
The x3750 M4 server can achieve almost linear performance scalability for HPC workloads going from eight to 16 to 32 cores. Also, IBM System x3750 M4 can provide more than four times improvement in HPC performance compared to the previous generation of x3850 X5 with Intel Xeon X7560 processors.
Based on the observed results, I can tell that x3750 M4 can potentially eliminate the need for multi-node HPC workload deployments or at least significantly reduce the number of nodes required to run workgroup HPC workload, making deployment and support much easier!
While IBM BladeCenter-based and IBM iDataPlex-based HPC cluster deployments are good for large scale implementations, the x3750 M4 is a reasonable choice for small single- or multi-node workgroup HPC clusters.
Ilya Krutov, IBM Redbooks