Summary
Our first real foray into workloads for scientific computing was a success on several levels. First, we had an opportunity to evaluate a server from Supermicro that left us quite impressed. The A+ Server 2021A-32R+F has an excellent balance of CPU power, memory capacity, direct attached storage and I/O capability in a 2U form factor with a highly efficient power supply. It is especially well suited for HPC applications and was highly appropriate for this review and workload exploration.
Since this review has taken us longer than normal to publish, the hardware platform is a little out-of-date at the moment. However, we’d like to point out the spiritual successor, the AS-2022G-URF. This newer version is based on AMD’s Fiorano platform, using the 8 or 12-core Magny Cours and four channels of DDR3 (upto 256GB) and is equally suitable for HPC applications.
Second, we gained experience with MAQSIP-RT, a prominent application in the weather forecasting and analysis community. MAQSIP-RT is a well tuned and highly scalable workload that shows good behavior from one to twelve CPUs. Finding scalable benchmarks is significantly challenging, as many applications have a variety of scaling bottlenecks. For example, even a database (a system with its own special language that is inherently designed to be parallel) such as MySQL runs into substantial scaling problems above 8 threads. Fundamentally, implementation details really matter when it comes to parallelism. This is the first real HPC application we have found for Linux and it does not disappoint at all. We hope that in the future, MAQSIP-RT will be a starting point of a full suite of benchmarks for our Linux based server reviews.
Again, we’d like to thank Carlie Coats from BaronAMS for helping us with MAQSIP-RT and Michael Kalodrich and Douglas Herz from Supermicro for providing us with the server.