Benchmarks
8 minute read
Network Performance
The Cruiser’s 2.5GbE uplink port provides substantial throughput improvements over standard Gigabit Ethernet:
| Metric | Result |
|---|---|
| Peak Upload | 2.35 Gbps |
| Peak Download | 1.93 Gbps |
| Bidirectional Total | 3.2 Gbps |
| CPU Usage | ~27% (single core) |
Test Configuration
Hardware
| Component | Details |
|---|---|
| Board | Cruiser Carrier Board |
| Compute Module | Raspberry Pi Compute Module 5 Lite Rev 1.0 |
| 2.5GbE Controller | Realtek RTL8156BG (USB 3.0) |
| Test Endpoint | MacBook Pro with 2.5GbE USB dock |
| Network | Direct connection via 2.5GbE switch |
Software
| Component | Version |
|---|---|
| OS | Debian GNU/Linux 13 (trixie) |
| Kernel | 6.12.47+rpt-rpi-2712 |
| iperf3 (Cruiser) | 3.18 |
| iperf3 (Client) | 3.20 |
Link Negotiation
Both endpoints confirmed at 2500 Mb/s full-duplex before testing:
Speed: 2500Mb/s
Duplex: Full
Link detected: yes
Benchmark Results
All tests ran for 30 seconds using TCP with default settings.
Single Stream Performance
| Direction | Throughput | Efficiency |
|---|---|---|
| Upload (TX) | 2.35 Gbps | 94% |
| Download (RX) | 1.93 Gbps | 77% |
Upload (client → Cruiser):
[ ID] Interval Transfer Bitrate
[ 7] 0.00-30.01 sec 8.22 GBytes 2.35 Gbits/sec sender
[ 7] 0.00-30.01 sec 8.22 GBytes 2.35 Gbits/sec receiver
Download (Cruiser → client):
[ ID] Interval Transfer Bitrate Retr
[ 7] 0.00-30.01 sec 6.75 GBytes 1.93 Gbits/sec 0 sender
[ 7] 0.00-30.01 sec 6.75 GBytes 1.93 Gbits/sec receiver
Multi-Stream Performance (4 Parallel)
| Direction | Throughput | vs Single Stream |
|---|---|---|
| Upload (TX) | 2.35 Gbps | Same |
| Download (RX) | 1.93 Gbps | Same |
Multi-stream tests confirm the link is fully saturated with a single stream — additional parallel connections do not increase throughput.
Bidirectional Performance
Simultaneous upload and download:
| Direction | Throughput |
|---|---|
| Upload (TX) | 2.34 Gbps |
| Download (RX) | 857 Mbps |
| Combined | 3.2 Gbps |
When contending for the USB 3.0 bus, upload traffic is prioritized. The combined bidirectional throughput exceeds 3 Gbps.
CPU Utilization
During maximum throughput testing:
| Metric | Value |
|---|---|
| iperf3 process | 27% (single core) |
| System overhead | ~10% (kernel/softirq) |
| CPU idle | 90% |
The 2.5GbE port leaves ample CPU headroom for other tasks.
Technical Notes
Why Download is Lower Than Upload
The 2.5GbE uplink port connects via USB 3.0, which has asymmetric performance characteristics. This is normal behavior for USB-based network adapters:
- Upload (host → device): USB bulk OUT transfers are highly efficient
- Download (device → host): USB bulk IN transfers have more protocol overhead
Despite this asymmetry, download throughput of 1.93 Gbps is still nearly double Gigabit Ethernet (1.0 Gbps).
Real-World Implications
For typical NVR/camera workloads where cameras upload footage to Cruiser:
- Camera streams: Upload direction benefits from full 2.35 Gbps
- Client viewing: Even reduced download speed (1.93 Gbps) vastly exceeds Gigabit
- 8x cameras: Each 1GbE camera can stream at full speed without an uplink bottleneck
Reproduce These Tests
Install iperf3
On Cruiser:
sudo apt update && sudo apt install -y iperf3
On test client (macOS):
brew install iperf3
Run Tests
Start server on Cruiser:
iperf3 -s
Run from client:
# Single stream upload
iperf3 -c CRUISER_IP -t 30
# Single stream download
iperf3 -c CRUISER_IP -t 30 -R
# Multi-stream upload (4 parallel)
iperf3 -c CRUISER_IP -t 30 -P 4
# Multi-stream download
iperf3 -c CRUISER_IP -t 30 -P 4 -R
# Bidirectional
iperf3 -c CRUISER_IP -t 30 --bidir
Verify Link Speed
Check link negotiation before testing:
# On Cruiser (eth1 is typically the 2.5GbE interface)
/usr/sbin/ethtool eth1 | grep Speed
Expected output: Speed: 2500Mb/s
Storage Performance
The Cruiser’s JMS561 SATA controller provides dual-port storage via USB 3.0:
| Metric | Single Drive | Dual Drive (Simultaneous) |
|---|---|---|
| Sequential Read | 168 MiB/s | 318 MiB/s |
| Sequential Write | 163 MiB/s | 301 MiB/s |
| Random 4K Read | 90 IOPS | — |
| Random 4K Write | 376 IOPS | — |
| CPU Usage | ~2% system | ~2% system |
Storage Test Configuration
| Component | Details |
|---|---|
| SATA Controller | JMicron JMS561 (USB 3.0 bridge, 2-port) |
| Test Drives | 2x WD Purple 2TB (WDC WD22PURZ-74B4ZY0) |
| Drive Type | 5400 RPM surveillance HDD |
| RAID Mode | JBOD (independent drives) |
| Benchmark Tool | fio 3.39 |
Sequential Performance
Both SATA ports deliver full HDD throughput. The JMS561 USB 3.0 interface (5 Gbps theoretical) does not bottleneck these drives.
Single drive:
Sequential Read: 168 MiB/s (176 MB/s)
Sequential Write: 163 MiB/s (171 MB/s)
Dual drive simultaneous I/O:
Sequential Read: 318 MiB/s (333 MB/s) — nearly 2x single drive
Sequential Write: 301 MiB/s (315 MB/s) — nearly 2x single drive
Both drives can operate at full speed simultaneously, confirming no controller bottleneck.
Random I/O Performance
Random 4K performance is HDD-limited (seek time), not interface-limited:
| Test | IOPS | Latency |
|---|---|---|
| Random 4K Read (QD=1) | 90 | 11ms |
| Random 4K Read (QD=32) | 90 | — |
| Random 4K Write (QD=1) | 376 | — |
Higher queue depth does not improve random read IOPS on HDDs — the mechanical seek time is the bottleneck. Random writes benefit from drive write caching.
RAID Modes
The JMS561 supports hardware RAID with four modes:
| Mode | Capacity | Description | Use Case |
|---|---|---|---|
| JBOD | 2x individual | Independent drives | Maximum flexibility |
| RAID 0 | Combined (2x) | Striping across drives | Maximum throughput |
| RAID 1 | Single drive | Mirroring for redundancy | Data protection |
| SPAN | Combined (2x) | Concatenation | Single large volume |
For configuration instructions, see Hardware Interface - RAID DIP Switch or Software - JMicron RAID Manager.
RAID Mode Performance Comparison
Benchmarked with 2x WD Purple 2TB HDDs (5400 RPM):
| Test | Single Drive | Both Drives (Parallel) | RAID 0 | RAID 1 |
|---|---|---|---|---|
| Sequential Read | 168 MiB/s | 318 MiB/s | 316 MiB/s | 168 MiB/s |
| Sequential Write | 163 MiB/s | 301 MiB/s | 311 MiB/s | 158 MiB/s |
| Random 4K Read | 90 IOPS | — | 112 IOPS | 119 IOPS |
| Random 4K Write | 376 IOPS | — | 149 IOPS | 367 IOPS |
“Both Drives (Parallel)” = JBOD mode with simultaneous I/O to both drives, simulating NVR workload (record + playback).
Key observations:
- RAID 0 nearly doubles sequential throughput (316/311 MiB/s vs 168/163 MiB/s single drive)
- RAID 0 random write IOPS drops significantly due to striping overhead on HDDs
- RAID 1 read performance equals single drive (no read striping in this controller)
- RAID 1 write performance slightly lower (writes go to both drives)
- Both drives in parallel achieves similar throughput to RAID 0 with more flexibility
Recommendations:
| Workload | Recommended Mode | Reason |
|---|---|---|
| Surveillance NVR | JBOD | Flexibility for separate record/playback |
| Maximum write speed | RAID 0 | ~2x sequential throughput |
| Critical data | RAID 1 | Automatic redundancy |
| Single large volume | SPAN | Simple combined capacity |
Hardware vs Software RAID Performance
Comparing the JMS561 hardware RAID against Linux mdadm software RAID:
RAID 0 (Striping):
| Test | Hardware (JMS561) | Software (mdadm) | Difference |
|---|---|---|---|
| Sequential Read | 316 MiB/s | 314 MiB/s | ~same |
| Sequential Write | 311 MiB/s | 302 MiB/s | ~same |
| Random 4K Read | 112 IOPS | 141 IOPS | mdadm +26% |
| Random 4K Write | 149 IOPS | 816 IOPS | mdadm +5.5x |
RAID 1 (Mirror):
| Test | Hardware (JMS561) | Software (mdadm) | Difference |
|---|---|---|---|
| Sequential Read | 168 MiB/s | 162 MiB/s | ~same |
| Sequential Write | 158 MiB/s | 155 MiB/s | ~same |
| Random 4K Read | 119 IOPS | 121 IOPS | ~same |
| Random 4K Write | 367 IOPS | 394 IOPS | mdadm +7% |
Key findings:
- Sequential throughput is identical — both are limited by USB 3.0 and drive speed
- Random IOPS significantly better with mdadm — especially RAID 0 writes (5.5x improvement)
- Hardware RAID’s striping overhead hurts small writes — the JMS561 adds latency for random I/O
- mdadm CPU usage is minimal — measured 3-5% during heavy writes (CM5 quad-core handles it easily)
See Software - Hardware vs Software RAID for a full comparison of when to use each approach.
SSD vs HDD Comparison
Testing with Samsung 870 EVO SSDs reveals the USB 3.0 interface limits:
| Test | HDD (5400 RPM) | SSD (870 EVO) | Bottleneck |
|---|---|---|---|
| Sequential Read | 168 MiB/s | 338 MiB/s | Drive (HDD) / USB (SSD) |
| Sequential Write | 163 MiB/s | 314 MiB/s | Drive (HDD) / USB (SSD) |
| Dual Drive Read | 318 MiB/s | 345 MiB/s | Drive (HDD) / USB (SSD) |
| Dual Drive Write | 301 MiB/s | 317 MiB/s | Drive (HDD) / USB (SSD) |
| Random 4K Read | 90 IOPS | 4,450 IOPS | Drive (HDD) / USB (SSD) |
| Random 4K Write | 376 IOPS | 6,198 IOPS | Drive (HDD) / USB (SSD) |
Key findings:
- USB 3.0 ceiling: ~340-350 MiB/s — a single SSD saturates the interface
- Dual SSDs gain almost nothing — bandwidth is shared, not doubled
- HDDs are drive-limited — USB 3.0 has headroom for 2x HDDs at full speed
- Random IOPS improved 16-49x with SSDs despite USB protocol overhead
For surveillance workloads, HDDs are the better choice:
- 5400 RPM drives fully utilize available bandwidth
- Lower cost per TB
- Better suited for continuous 24/7 streaming writes
- SSDs provide no sequential throughput benefit with dual drives
Real-World Implications
For NVR/surveillance workloads:
- 8x 1080p cameras @ 8Mbps: 8 MB/s total — easily handled by single drive
- Recording + playback: Dual drives in JBOD allow separate record/playback operations
- RAID 1 mirroring: Protects against single drive failure with no performance penalty for writes
Reproduce Storage Tests
Install fio:
sudo apt update && sudo apt install -y fio
Run benchmarks:
# Sequential write (1M blocks, 60 seconds)
sudo fio --name=seq_write --filename=/dev/sda --direct=1 --rw=write \
--bs=1M --size=4G --numjobs=1 --runtime=60 --time_based --group_reporting
# Sequential read
sudo fio --name=seq_read --filename=/dev/sda --direct=1 --rw=read \
--bs=1M --size=4G --numjobs=1 --runtime=60 --time_based --group_reporting
# Random 4K read
sudo fio --name=rand_read --filename=/dev/sda --direct=1 --rw=randread \
--bs=4k --size=1G --numjobs=1 --iodepth=1 --runtime=60 --time_based --group_reporting
# Dual drive simultaneous read
sudo fio --name=dual_read --filename=/dev/sda:/dev/sdb --direct=1 --rw=read \
--bs=1M --size=4G --numjobs=1 --runtime=60 --time_based --group_reporting
These tests write directly to the raw block device. All data on the drive will be destroyed. Only use on drives you can wipe.
Test Date
Benchmarks performed: January-February 2026
Last modified October 23, 2026