Skip to main content

Performance Characteristics

Metadata Performance​

In traditional file systems, metadata operations (create, stat, delete, rename) incur millisecond-scale latency due to synchronous disk I/O and global locks.

MASS maintains the metadata index in a high-speed storage tier backed by DRAM and NVMe. Metadata I/O is served immediately from the DRAM write buffer, achieving latencies in the hundreds of nanoseconds to low microseconds range. With no global metadata lock and independent per-target processing, the metadata server bottleneck present in traditional architectures is structurally absent.

Concurrent I/O Scalability​

Each I/O target has an independent SPDK I/O queue and is pinned 1:1 to a CPU core. With no serialization points between targets, throughput is maintained as client count increases.

In traditional distributed file systems, hundreds of clients concurrently accessing a metadata server causes lock contention that saturates throughput. MASS eliminates this bottleneck by design.

HPC / MPI Workloads​

In MPI-IO workloads where thousands of processes concurrently access a single shared file, traditional systems suffer severe performance degradation from POSIX byte-range lock contention.

MASS supports concurrent writes to Array Objects:

  • Thousands of clients can concurrently write to different regions of the same Array Object.
  • Independent processing per object and byte range means no inter-process locking.
  • MPI-IO, HDF5, and POSIX layouts all operate on this model.

Linear Scaling​

Adding storage nodes increases the number of I/O targets linearly. Because data placement is determined by client-side deterministic computation, new targets immediately begin serving I/O with no central bottleneck.

Under ideal conditions, the throughput limit of MASS is determined by NVMe device and fabric network bandwidth — not software bottlenecks.