18:20 - 19:00
A misunderstanding of storage system characteristics is often the reason behind an inability to nicely scale critical jobs to a required level of performance – be it in DeepLearning, HPC or other areas, on-prem or in the cloud. Elbencho to the rescue! To make it easy to characterize or validate a modern storage system for a certain workload, Sven Breuner, the creator of the BeeGFS parallel file system and nowadays Field CTO at VAST Data, has published elbencho. Elbencho is a new vendor-neutral storage benchmark that can simulate a wide variety of access patterns, ranging from lots of small files over random IOPS in large shared files up to low level testing of underlying block devices. For DeepLearning, elbencho can even include GPUs in the data access. Elbencho is available on github for everyone to use and contribute.