10-client IO-500 numbers are apples vs. oranges
It isn't clear why you continue to tout the IO-500 10-client storage results. Weka used NVMe PCI storage, Summit used NLSAS HDDs. Half of the IO bandwidth of the result is based on IOPS, so there is no question that NVMe is going to beat HDDs on this workload. How is the scalability of Weka vs. GPFS or Lustre? There are no results on the IO-500 with more than 17 clients.