At Oracle, we work closely with independent software vendors (ISVs) to ensure that they get the best performance for their products on our cloud. On Oracle Cloud Infrastructure (OCI), ISVs can get unmatched speed, security, and scalability at the best price per performance in the market. To learn more about Oracle’s ISV network, visit ISVs bet their business on Oracle Cloud.

We recently worked with CRAFT Tech, a company that specializes in computational fluid dynamics (CFD) simulations, to benchmark their application CRUNCH CFD. CRUNCH CFD is an unstructured solver with Multiphysics simulation abilities that allows you to analyze complex flow problems in a commutable framework, with specialized fluid and thermal solvers. To learn more about this application, visit their website.

Benchmark setup

We created a cluster of BM.Optimized3.36, Oracle’s latest high-performance computing (HPC) instances, using our HPC Cluster Marketplace stack. This stack allows you to easily deploy a cluster with customizable features, such as the ability to add network file share (NFS) share volumes, and install SLURM. With the ease of usability, these stacks are up within minutes of launching the stack, reducing the time to get workloads up and running.

The following graphic shows the system specs of our benchmark setup and that of AWS:

A chart showing the comparison between OCI HPC and AWS with and without EFA.

Both AWS systems used c5n.18xlarge, but AWS with EFA had Elastic Fabric Adapter (EFA) as the interconnect. EFA is a network device that enables applications to communicate directly with the network interface hardware. In comparison, the Oracle setup used InfiniBand RDMA, which provides low-latency networking (as low as 1.7 µs) and scalability over thousands of cores. These setups also have no added overhead on the network to limit throughput.

CRUNCH CFD result

We used the 9-Species Hydrogen-Oxygen Petersen & Hansen Reaction Mechanism with k-ε Turbulence Model in three different cases: 5 Million cell, 15 Million cell, and 44 Million cell mesh size. The metrics taken were total execution time, time per iteration, and total time processing. The following graphic describes each metric:

A graphic depicting the total time execution, time per iteration, and total preprocessing time.

The following graph shows the comparison between BM.Optimized3.36 and c5n.18xlarge with EFA for total preprocessing time, where the lower time is better:


Figure 1: Parallel preprocessing time: Grid and solution partitioning comparison between OCI HPC and AWS with EFA. OCI HPC is up to 70% faster than AWS with EFA.

The following graph shows the comparison between BM.Optimized3.36, c5n.18xlarge with 100 GB/s Ethernet, and c5n.18xlarge with EFA for time per iteration, where the lower time is better:


Figure 2: The CPU time per iteration in the steady state reaction between OCI HPC, AWS with EFA, and AWS without EFA. OCI HPC is 18% faster than AWS without EFA and 10% faster than AWS with EFA.

The following graph shows the comparison between BM.Optimized3.36, c5n.18xlarge with 100 GB/s Ethernet, and c5n.18xlarge with EFA for total time run, where the lower time is better.


Figure 3: The total time of the steady state reaction between OCI HPC, AWS with EFA, and AWS without EFA. OCI HPC is 22% faster than AWS without EFA and 20% faster than AWS with EFA.

The performance increase directly translates to more cost savings. Let’s use the example of the 44-M cell model to demonstrate. The cost per job for c5n.18xlarge with EFA ranges from $7.00 at 256 cores to $10.00 at 1,024 cores while the cost per job for BM.Optimized3.36 is $4.00 at 256 cores to $5.00 at 972 cores. So not only are we getting shorter run times, the BM.Optimized3.36 costs less per hour at $2.70 compared to that of c5n.18xlarge at $3.88. With around twice the cost savings on OCI, you don’t need to run CRUNCH CFD anywhere else.

Conclusion

Overall, Oracle’s HPC cluster shows clear performance advantage over AWS, with 15% improvement in time and iteration. Oracle’s NVMe is also more than 50% faster I/O speed for parallel file partitioning than AWS SSD.

Besides CRUNCH CFD, other CFD applications are running faster than ever on OCI. If you’d like to test these applications yourself, start your 30-day free trial and get access to a wide range of Oracle Cloud Infrastructure services for 30 days.