storage
arrow_backBack to Blog
Engineering|20 MIN READ|FEB 22, 2026

NVMe RAID Tuning for High-Throughput Chains

Benchmarking various disk configurations to minimize I/O wait times during heavy state sync periods.

terminal

01node Team

Infrastructure Engineers

Storage performance is the most underappreciated bottleneck in validator operations. While operators obsess over CPU cores and RAM capacity, it is often disk I/O that determines whether your validator keeps up with a high-throughput chain during peak load. This article documents our extensive benchmarking of NVMe RAID configurations for blockchain workloads.

The Storage Challenge

Modern high-throughput chains generate enormous storage demands. A Solana validator writes 50–100 GB of ledger data per day. An Ethereum archive node exceeds 15 TB. Sui’s object-based storage model creates millions of small random reads per second during state queries.

These workloads are fundamentally different from traditional database or web server I/O patterns. Blockchain storage is characterized by: - Heavy sequential writes during block processing and state commitment. - Intense random reads during state verification and RPC query serving. - Periodic bulk reads during state sync and snapshot restoration. - Write amplification from LSM-tree based storage engines (RocksDB, PebbleDB).

Test Methodology

We benchmarked five NVMe configurations using identical Samsung PM9A3 3.84TB drives on AMD EPYC 7763 platforms:

1. Single NVMe: Baseline, no redundancy. 2. RAID 0 (2x NVMe): Striping for maximum throughput. 3. RAID 1 (2x NVMe): Mirroring for redundancy. 4. RAID 10 (4x NVMe): Striped mirrors for throughput + redundancy. 5. ZFS mirror (2x NVMe): Software RAID with compression and snapshots.

Each configuration was tested against three real-world blockchain workloads: - Solana validator: 72-hour production run measuring vote latency and slot processing time. - Cosmos SDK state sync: Full state sync of the Cosmos Hub from genesis. - RPC query load: 10,000 concurrent RPC requests simulating production dApp traffic.

Key Findings

RAID 10 delivered the best overall performance for mixed validator workloads, with 2.3x sequential write throughput and 1.8x random read IOPS compared to a single drive, while maintaining full redundancy.
ZFS mirror surprised us with competitive performance numbers. The inline LZ4 compression reduced effective write volume by 30–40% for blockchain data (which compresses well due to repetitive data structures). The snapshot capability also provides zero-downtime backup for validator state.
RAID 0 is a trap. While it shows the best synthetic benchmarks, losing a drive means losing the entire validator state. For chains with multi-day state sync times, this translates to potentially a week of downtime—and missed rewards.

Tuning Parameters That Matter

Beyond RAID level selection, these kernel and filesystem parameters had the largest impact on validator performance:

I/O Scheduler: `none` (or `noop`) for NVMe devices. The default `mq-deadline` scheduler adds unnecessary overhead for devices with native command queuing.
Filesystem: XFS outperformed ext4 by 15–20% on blockchain write workloads. For ZFS, `recordsize=16K` matched the typical RocksDB SST block size.
Read-ahead: Reducing `read_ahead_kb` from the default 128 to 32 improved random read performance by 25%. Blockchain random reads are small and unpredictable; large read-ahead buffers waste I/O bandwidth.
Swap: Disabled entirely. Validator processes that swap to disk will miss blocks. If you are swapping, you need more RAM, not faster storage.

Our Production Configuration

Based on these benchmarks, our current production standard for high-throughput validators (Solana, Sui, Aptos) is:

- 4x Samsung PM9A3 3.84TB in RAID 10 via mdadm. - XFS filesystem with noatime,nodiratime,discard mount options. - I/O scheduler set to none. - read_ahead_kb=32. - Dedicated NVMe for operating system and logs (separate from validator data).

For Cosmos SDK chains and lower-throughput validators, we use ZFS mirrors with LZ4 compression, trading marginal performance for operational convenience (snapshots, compression, self-healing).

Monitoring Storage Health

NVMe drives have finite write endurance. Our eBPF monitoring stack (see our article on zero-latency monitoring) tracks per-drive SMART data including: - Percentage of drive life used. - Uncorrectable error counts. - Temperature and thermal throttling events.

We replace drives proactively at 80% endurance consumed, well before failure risk increases. For our highest-throughput validators, this means drive replacement every 18–24 months.

StakingInfrastructureValidatorEngineering01node

Share this article

Help others discover quality infrastructure insights.

ShareLinkedIn
mail

Stay updated

Follow us for more technical insights, infrastructure deep-dives, and ecosystem updates.