Production Hardware Guide

Planning considerationsCopy Anchor Link

Before deploying RustFS in production, assess your workload and capacity requirements over a two- to three-week period. Focus on the following factors:

1. Data volume

   • Initial data volume: Estimate the usable data volume at go-live (in TiB), including the ratio of hot to cold data.
   • Growth forecast: Estimate data growth for the next 24 months based on business plans. We recommend modeling growth quarterly.
   • Object count: Estimate the total number of objects from the average object size (ideally 128 KB to 1 MB). Deployments with more than 100 million objects need additional tuning.

2. Workload characteristics

   • Access patterns: Distinguish read-intensive workloads (such as content distribution) from write-intensive workloads (such as log collection).
   • Compliance requirements: Retention periods must meet industry regulations (for example, some financial regulations require at least five years of retention).
   • Multi-site deployment: For cross-region deployments, assess network latency (ideally under 50 ms) and bandwidth costs.

3. Storage architecture

   • Bucket planning: Organize buckets by business unit. We recommend no more than 500 active buckets per cluster.
   • Disaster recovery: Choose active-active replication (recommended) or asynchronous replication based on how critical the data is.

Hardware Configuration MatrixCopy Anchor Link

Baseline configurations based on load-test results:

Deployment guidelines:

1. Use homogeneous nodes: every node should have identical hardware from the same batch and the same firmware versions.
2. The network must use a leaf-spine topology, a physically isolated storage network, and dual uplinks.
3. We recommend 2U servers with at least 12 drive bays per node, depending on the number of drives required.

Performance tuningCopy Anchor Link

Network Topology Optimization (Highest Priority)Copy Anchor Link

• Bandwidth sizing: Provision 0.5 Gbps of bandwidth per TB of usable data (for example, 100 TB of data needs 50 Gbps of dedicated bandwidth).
• Latency Requirements:
  • Inter-node P99 latency ≤ 2ms
  • Cross-rack latency ≤ 5ms

Storage Subsystem TuningCopy Anchor Link

• Controller Configuration:
  • Enable read-ahead cache (recommended 256 MB+)
  • Disable all RAID features and use pass-through (HBA/JBOD) mode.
  • Check BBU health regularly.
• SSD Parameters:
  • Reserve 20% over-provisioning (OP) space to improve endurance.
  • Enable atomic writes (requires hardware support).

Memory allocationCopy Anchor Link

• Allocation Ratios:
  • Metadata cache: 60% of total memory
  • Read/write buffers: 30%
  • System reserve: 10%

Network Design Reference ModelCopy Anchor Link

Network bandwidth vs. disk countCopy Anchor Link

Example: A video platform runs a 16-node cluster. Each node has:

• 8×7.68 TB NVMe SSD
• Dual 100GbE NVIDIA ConnectX-5 NICs
• Achieves aggregate throughput of 38 GB/s

Memory sizingCopy Anchor Link

Use the following formula to size memory based on disk capacity and workload type:

# Memory calculation formula (unit: GB)
def calc_memory(data_tb, access_pattern):
 base = 32 # Base memory
 if access_pattern == "read_heavy":
 return base + data_tb * 0.8
 elif access_pattern == "write_heavy":
 return base + data_tb * 1.2
 else: # mixed
 return base + data_tb * 1.0

Reference sizes:

Storage guidelinesCopy Anchor Link

Choosing storage mediaCopy Anchor Link

File System ConfigurationCopy Anchor Link

# XFS formatting example
mkfs.xfs -f -L rustfs_disk1 -d su=256k,sw=10 /dev/sdb

# Recommended mount parameters
UUID=xxxx /mnt/disk1 xfs defaults,noatime,nodiratime,logbsize=256k 0 0

High availabilityCopy Anchor Link

1. Power Supply:

   • Use dual (redundant) power supplies.
   • Connect each PDU to a separate power feed.
   • Provide a UPS with at least 30 minutes of runtime.

2. Cooling Requirements:

   • Rack power density ≤ 15 kW per rack.
   • Keep the inlet-to-outlet temperature difference (ΔT) within 8 °C.

3. Firmware Management:

   • Maintain a hardware compatibility matrix.
   • Standardize on the same firmware versions across nodes.

This guide is based on the latest RustFS development version. For production deployments, fine-tune these settings using your hardware vendor's documentation. For quarterly hardware health assessments, contact RustFS.