Data center storage insights center

Characterizing the storage workload for AI training systems faces two unique challenges that the MLPerf Storage Benchmark Suite aims to address: the cost of AI accelerators and the small size of available datasets. This blog shows how the MLPerf Storage benchmark addresses both.

MLPerf Storage tool is extremely helpful in benchmarking storage for various models by reproducing realistic AI workloads. Read this blog to learn how Micron is using MLPerf to test storage for AI workloads.

When running and collecting workload traces of RocksDB, we sometimes see large latency spikes. In this blog, we talk about the methods used to identify the root cause of latency spikes in a mixed read and write workload. 

Google and Meta have worked closely to introduce the Flexible Data Placement (FDP) mode into the NVMe spec. In this blog, our testing shows that FDP decreases write amplification by 60% for sequential workloads.

db_bench is Meta’s preferred workload testing methodology as it provides a good emulation of workloads at the query level and simulates precise RocksDB storage I/Os. In this blog, we talk about a test that compares the performance of Micron's 7450 vs Samsung's PM9A3 vs Solidigm's D5-P5430 for RocksDB.

Read about our deeper dive into the Apache Cassandra workload featured in our recently published tech brief comparing 6500 ION performance to a competitor's QLC drive. Bursty workloads require SSDs that perform well at high average disk IOs — an area where the Micron 6500 ION excels.

Check out these test results for high-performance computing (HPC) and AI using WEKA Data Platform software. We coupled this software with Supermicro servers built on the 4th Gen AMD EPYC™ 9554 and the Micron 6500 ION SSD.

In this blog, Ceph object storage performance with Micron 6500 ION SSDs is compared to hard disk drives (HDDs). Read this article and see how the Micron 6500 ION wins in performance, power and cost over HDDs.