ClusterFS2 NFS Latency Testing and Benchmarking
Benchmark Methodology
ClusterFS2 NFS mounts were tested across all lab.franklin compute nodes using a standardized seq/rd test pattern with 1MB blocks, measuring read latency (mean, p95, p99), write throughput (MB/s), and I/O scheduler behavior under concurrent access. Tests ran for 30-minute windows per host.
Results Summary
| Host | Read Latency (ms) | Write Throughput (MB/s) | Notes |
|---|---|---|---|
| skynet (10.10.16.10) | 0.42 / 0.87 / 1.31 | 248.5 | SATA SSD-backed NFS export |
| chonk (10.10.11.5) | 0.31 / 0.64 / 0.98 | 312.1 | NVMe-backed, optimal mount options |
| stargate (10.10.15.2) | 0.55 / 1.12 / 1.87 | 195.3 | Default mount — no tuning applied |
| node3 (10.10.12.3) | 0.48 / 0.93 / 1.45 | 223.7 | Ethernet-limited, no jumbo frames |
Mount Option Tuning Results
Applied the following mount options to stargate's clusterfs2 workspace mount:
rsize=1048576,wsize=1048576— increased read/write buffer sizes from 8KB defaults to 1MB, matching the benchmark block size.noatime,nodiratime— eliminated unnecessary metadata updates on read-only workloads (Ansible role collection reads).soft,retrans=3,timeo=600— replaced hard mount with soft for non-critical paths, reducing stall time on transient NFS server unavailability.
Post-tuning write throughput on stargate improved from 195.3 MB/s to 287.4 MB/s (+47%), bringing it within 8% of chonk's optimal NVMe baseline. Further gains require jumbo frame (MTU 9000) configuration across the switch fabric.