Storage performance is critical for large-scale simulation, big data analysis, and artificial intelligence (AI). The IO-500 list serves as a resource for companies looking to improve their storage performance and it has become particularly critical with the explosion of GPU based workloads in the enterprise. Each system is evaluated by the IO-500 benchmark that measures the storage performance using read/write bandwidth for large files and read/write/listing performance for small files.
WekaIO recently submitted results to the IO-500 and came within 5% of the world’s biggest supercomputer storage installation powered by IBM Spectrum Scale. In the ten node challenge – a benchmark that measures how well the storage system can satisfy a small cluster of 10 compute clients – WekaIO came within 5% of the world’s biggest supercomputer, Oak Ridge Summit.
This was an incredible accomplishment as it reinforces our claim that we are the world’s fastest file system for AI and technical HPC workloads. Given that we are sitting in #2 position, I think it is important to clarify that our claim relates to how efficiently we deliver I/O from the underlying hardware, not how big our hardware footprint is. If you throw enough hardware at any storage problem it will eventually improve performance. Therefore it is important to differentiate between biggest and fastest.
You see, the IBM Spectrum Scale storage solution at Oak Ridge Summit can lay claim to being the world’s biggest storage system. Let’s start with the basics, the IBM ESS Storage system took up 40 racks of space while WekaIO storage fitted into less than half a rack of storage servers.
However if you look at the infrastructure used to drive performance it is clear that WekaIO is the performance leader. The IBM system is comprised of 77 storage servers compared to WekaIO’s 32. The IBM system was populated with 1 Terabyte of memory per storage server compared to just 192GBytes of memory for WekaIO. In addition each IBM storage node had 8 NVMe drives to boost small file I/O performance while WekaIO had only 6 NVMe drives per node.
In summary, it took 2.4x more server nodes, 12x more memory and 3.2x more NVMe for IBM Spectrum Scale to deliver 5% more performance than the WekaIO Matrix File system.
If you look at the performance score per storage node, WekaIO is over double the performance of IBM.
|Number of Nodes||77||32||141%|
|Total Memory (GB)||77,000||6,144||1153%|
|Number of NVMe drives||616||192||221%|
To check out the specifications of the IBM Summit Storage System, click on the link below.