RandRW Benchmark
BrewFS delivers 92% more fully drained mixed random-I/O throughput versus JuiceFS in the July 24, 2026 Redis + RustFS snapshot (218.60 vs 113.86 MiB/s).
Published for reproducibility
Every value below is part of the benchmark contract. This snapshot uses the published runner profile and records its effective parameters with the raw artifacts.
Metrics captured: IOPS, throughput, average latency, P95, and P99 — for both cold-cache and warm-cache states.
| Benchmark Type | RandRW |
| Comparison | BrewFS vs JuiceFS |
| Workload Model | Agent-style random read/write |
| Filesystem Access | FUSE |
| Data Backend | RustFS, S3-compatible |
| Metadata Backend | Redis |
| Block Size | 4 MiB |
| Concurrency | 8 fio jobs, iodepth=1 |
| Per-job size / runtime | 512 MiB / 20 s (time-based) |
| Read / Write Ratio | 70 / 30 |
| Cache State | Durable prefill, drain, cache-preserving remount |
| Benchmark Tool | fio with io_uring |
| Mount Profile | writeback-throughput-profile |
Recommended fio profile
This fio shape matches the published mixed random-I/O workload; use the repository runner to reproduce mount, prefill, drain, and cache handling.
[agent-randrw]
rw=randrw
rwmixread=70
bs=4m
iodepth=1
numjobs=8
size=512m
runtime=20
time_based=1
direct=0
group_reporting=1What the benchmark measures
Overall RandRW Performance
Relative throughput, higher is better
IOPS & Throughput
Mixed 70/30 read/write, relative
Latency (lower is better)
avg · P95 · P99
Multi-agent Scaling
Aggregate throughput by concurrent agents
The fully drained comparison is the published result. IOPS, latency, and multi-agent scaling panels remain illustrative until matching results are published.
RandRW is a better proxy for agent workloads
Agents continuously mix small reads, partial writes, metadata access, and artifact creation — not pure sequential read/write. BrewFS is optimized for this shared, random, and concurrent filesystem access.