Performance Characteristics

MASS distributes data processing across storage nodes and NVMe devices. The performance an application sees depends on how its requests use those resources, as well as the client and network configuration.

Parallel Access

Requests to different data can be processed in parallel across nodes and devices. Multiple clients can make use of this parallelism, provided their access pattern distributes the work.

When requests concentrate on the same data or device, the available parallelism is lower. Increasing client count alone does not guarantee higher throughput.

Metadata and Small I/O

File creation, directory traversal, and small reads and writes involve more operations per byte than large sequential transfers. Their performance depends on request latency, client processing, and how frequently clients access shared metadata.

A bandwidth measurement from large transfers therefore does not describe small-file or metadata-heavy behavior. Evaluate the operations and file sizes that the application actually uses.

HPC and Shared Files

Parallel applications can access different regions of a shared dataset concurrently. File layout, request size, synchronization, and the application's MPI-IO or HDF5 integration influence how effectively those requests are distributed.

Concurrent writes to overlapping data still require coordination by the application or its I/O library. Parallel storage access does not remove those consistency requirements.

Scaling and Shared Resources

Adding storage nodes increases available capacity and can add processing and I/O bandwidth. The gain depends on data placement, workload parallelism, client capability, and network bandwidth.

Data protection, recovery, and other workloads consume resources too. File and object I/O and KV cache I/O use different software paths, but can share physical resources such as network links. Assess throughput and latency under the mix of workloads expected in operation.