VPS Storage Performance: How to Benchmark and Optimize Disk I/O on Your Server

Why Disk I/O Performance Matters on a VPS

Storage I/O is often the most contended resource on a virtual private server. While CPU and RAM are typically dedicated to your instance, disk I/O is shared across the hypervisor’s storage subsystem. Poor I/O performance leads to slow database queries, laggy web applications, and extended deployment times. This guide walks you through benchmarking your VPS disk I/O with industry-standard tools, interpreting the results, and applying optimizations that make a measurable difference.

The first step in any optimization journey is understanding your baseline. The quality of underlying storage varies significantly between VPS providers — those using NVMe with low oversubscription deliver dramatically better I/O than older SATA SSD or HDD setups. Compare providers with transparent storage specs at virtualserversvps.com.

Benchmarking Tools: fio, hdparm, and dd

1. fio — The Gold Standard

fio (Flexible I/O Tester) is the most comprehensive disk benchmarking tool available for Linux. It supports customizable I/O patterns, queue depths, block sizes, and concurrency. Install it with sudo apt install fio and run these essential tests:

Random 4K reads (IOPS test):
fio --name=randread --ioengine=libaio --iodepth=32 --rw=randread --bs=4k --direct=1 --size=1G --numjobs=4 --runtime=60 --group_reporting

Random 4K writes (IOPS test):
fio --name=randwrite --ioengine=libaio --iodepth=32 --rw=randwrite --bs=4k --direct=1 --size=1G --numjobs=4 --runtime=60 --group_reporting

Sequential 1M reads (throughput test):
fio --name=seqread --ioengine=libaio --iodepth=16 --rw=read --bs=1m --direct=1 --size=4G --numjobs=2 --runtime=60 --group_reporting

Sequential 1M writes (throughput test):
fio --name=seqwrite --ioengine=libaio --iodepth=16 --rw=write --bs=1m --direct=1 --size=4G --numjobs=2 --runtime=60 --group_reporting

Interpreting fio Results

Metric	Good (NVMe)	Fair (SATA SSD)	Poor (HDD)
4K random read IOPS	>50,000	10,000–50,000	<500
4K random write IOPS	>30,000	5,000–30,000	<300
1M sequential read (MB/s)	>1,000	300–1,000	<150
1M sequential write (MB/s)	>500	200–500	<100
Latency (4K random, µs)	<200	200–1,000	>5,000

2. hdparm — Quick Cache/Transfer Tests

hdparm provides a quick buffered and cached read speed test. It’s useful for a rapid sanity check but doesn’t measure random I/O or write performance, so take results with a grain of salt:

sudo hdparm -Tt /dev/vda

The -T flag tests cached reads (RAM speed), and -t tests buffered disk reads. On NVMe VPS, expect -t results above 500 MB/s. Values below 200 MB/s suggest SATA SSD or significant hypervisor contention.

3. dd — Simple Sequential Test

The classic dd command can give a rough sequential throughput estimate. While not as precise as fio, it’s available on every Linux system:

Write test:
dd if=/dev/zero of=./test bs=1M count=4096 conv=fdatasync status=progress

Read test:
dd if=./test of=/dev/null bs=1M status=progress

Note: dd only measures sequential access. A provider with fast sequential reads but terrible random IOPS (common on shared HDD arrays) will look good in dd benchmarks but perform poorly under real database workloads.

Optimizing Disk I/O on Your VPS

1. Choose the Right I/O Scheduler

The I/O scheduler determines how the kernel orders and batches disk requests. For NVMe drives, the none scheduler (multi-queue block layer) offers the lowest latency. For SATA SSDs, mq-deadline provides a good balance. Check your current scheduler with cat /sys/block/vda/queue/scheduler and change it temporarily with echo none | sudo tee /sys/block/vda/queue/scheduler. Make it permanent via GRUB_CMDLINE_LINUX="elevator=none" in /etc/default/grub.

2. Filesystem Selection and Mount Options

ext4 is the default choice for most Linux VPS images. It’s stable, widely supported, and offers good performance with noatime mount option to disable access time updates. Add defaults,noatime to your /etc/fstab mount options.

XFS excels at large files and parallel I/O. It’s the default on RHEL/CentOS 8+ and offers better performance for workloads with many concurrent writers. Use mount options defaults,noatime,allocsize=1m for database workloads.

btrfs with Zstd compression can improve effective throughput for compressible data (text files, logs, source code). Use compress=zstd:3 mount option. However, btrfs has higher CPU overhead and is less battle-tested than ext4/XFS on VPS.

3. Tune the Block Layer

Several sysctl parameters affect I/O performance. The most impactful are:

• vm.dirty_ratio — Set to 20 (default 30) to reduce writeback buffering
• vm.dirty_background_ratio — Set to 5 (default 10) to start writeback earlier
• vm.dirty_expire_centisecs — Set to 3000 (default 30000) to flush dirty pages faster
• vm.vfs_cache_pressure — Set to 200 (default 100) to reclaim dentry/inode caches more aggressively

4. Database-Specific Optimizations

If your VPS runs MySQL or PostgreSQL, these additional I/O optimizations apply:

• Set innodb_flush_method=O_DIRECT (MySQL) to bypass the page cache
• Increase innodb_io_capacity to 2000+ for NVMe (default 200)
• Set effective_io_concurrency=200 (PostgreSQL) for NVMe
• Place database WAL/logs on a separate disk if possible
• Use discard=async mount option instead of periodic fstrim for TRIM

5. Monitor I/O Pressure

Use iostat -x 1 to monitor await (average I/O response time) and %util (percentage of time the device was busy). Values above 10ms await or 90% util indicate I/O saturation. For deeper insight, use /proc/pressure/io (Pressure Stall Information) which shows the percentage of time processes were stalled on I/O:

cat /proc/pressure/io — look at the full line. If avg60 exceeds 10%, your system is under significant I/O pressure.

Conclusion

Benchmarking your VPS disk I/O with fio gives you an objective baseline, while optimizations like the right I/O scheduler, filesystem choices, and kernel parameters can improve real-world performance by 20-40%. Remember that the most impactful factor is the underlying hardware — a VPS on NVMe with dedicated resources will always outperform shared SATA storage regardless of software tuning. For providers that offer transparent storage specifications, compare options at virtualserversvps.com.