Skip to content

How Fast Is Your vServer's Disk? What We Measured and Improved

We benchmarked the disks of our vServers, on local NVMe and in our Ceph cluster. The biggest bottlenecks were a tight limit and a small cache. Here are the numbers before and after.

Eric10 min18 November 202645 Views
How Fast Is Your vServer's Disk? What We Measured and Improved

How fast is your vServer's disk really, and where does it slow down? In November 2026 we stopped guessing and measured it: with test VMs directly on our hosts, with real customer load over several days, and with every setting changed on its own and measured before and after. Here are the numbers and what we changed as a result.

Key takeaways

  • More throughput for every disk: The limit per virtual disk goes from 200 to 1,000 MB/s read and from 150 to 1,000 MB/s write. On our NVMe hosts in the US, a VM now writes at 1,003 instead of 158 MB/s.
  • Ceph in Eygelshoven: A larger write cache raises sequential writes per disk from just over 410 to 1,000 MB/s. Without the limit it would be almost 2 GB/s.
  • Small writes stay fast: On Ceph, the cache acknowledges a single 4K write in 0.07 ms instead of around 2 ms.
  • Fewer dropped packets in the storage network: New buffer settings on the switches cut drops on the busiest link by 72%.
  • New for databases: In the ZapX panel you can now set the disk cache to "Database-optimized". On local NVMe that means up to 60% more committed writes.
  • What you need to do: If your vServer runs on Ceph (Eygelshoven location), restart it once in the web interface. Rebooting from inside the server is not enough.

Two kinds of storage: local and Ceph

Your vServer disk lives on one of two systems:

  • Local NVMe storage (Dallas, Ashburn, Los Angeles): The data sits on SSDs inside the host, combined into a RAID. Short paths, very low latency.
  • Ceph (Eygelshoven): A dedicated storage cluster with 360 NVMe SSDs. Every block of your disk is stored three times on different servers. If a host fails, your data is still there, and we can move VMs while they are running.

Both systems are fast, but in different ways. That is why we measured them separately.

How we measured

For each test we started a dedicated VM on a real host: Debian 13, 4 vCPU, 4 GB RAM, disks configured like customer disks. We measured with fio, direct I/O without caching in the guest, 30 to 60 seconds per test. We changed every setting on its own and measured before and after, repeating noisy values in several rounds. On top of that we analysed the real load of all customer VMs: on two comparable hosts for three days in 5-minute steps, and on Ceph for 30 hours in 15-minute steps.

Throughput of one virtual disk in MB/s, measured inside a test VM. Grey: before, green: after. ↑ write, ↓ read. Green is the new limit of 1,000 MB/s in each case; on Ceph the disk reached 1.9 to 2.0 GB/s without it (chart: ZAP-Hosting)

Local NVMe storage: the limit was the bottleneck

Customer disks always showed the same values: 158 MB/s write, 210 MB/s read. That was not the hardware, it was the per-disk limit of 150 and 200 MB/s. Without the limit, the same VM does many times more:

Test inside the VMwith old limitwithout limit
Sequential write (1 MB blocks)158 MB/s3,245–4,727 MB/s
Sequential read (1 MB blocks)210 MB/s9,996–11,446 MB/s
Random write (4K, parallel)38,500 IOPS57,000–64,000 IOPS
Random read (4K, parallel)51,000 IOPS83,000–88,000 IOPS

We still need a limit: many VMs share the SSDs of one host, and a single VM must not be able to slow down the others. But the real load shows how much headroom there is. All VMs on a host together wrote at a median of 52 MB/s, and the SSDs peaked at around 8,000 writes per second. The same SSDs handle more than 200,000 in our tests.

So in November we raised the limit to 1,000 MB/s read and 1,000 MB/s write per disk. We did it live, without a restart. Inside the VM we now measure 1,003 MB/s in both directions. For small accesses the limit hardly matters anymore: 60,000 random 4K writes per second are only about 240 MB/s.

RAID5 or RAID10?

In the US we are gradually switching hosts with four NVMe SSDs from RAID10 to RAID5. On the same hardware, RAID5 gives you 50% more usable space. The price is higher write latency, because every write also recalculates parity. We wanted to know whether you would notice.

Test inside the VM, without limitRAID5RAID10
Sequential write (parallel)3,245 MB/s4,727 MB/s
Sequential write (single stream)639 MB/s1,349 MB/s
Sequential read9,996 MB/s11,446 MB/s
Random write (4K, parallel)64,000 IOPS57,000 IOPS
Single writes with fsync (like databases)2,840/s, 0.34 ms10,516/s, 0.08 ms

With the 1,000 MB/s limit, both reach the limit for parallel writes and for reads, so you will not notice a difference there. If a single process writes continuously, RAID5 manages about 640 MB/s while RAID10 reaches the limit. RAID5 is most noticeable for applications that wait for confirmation after every single write, such as databases that fsync after each transaction. Under real customer load over three days, the median write latency was 0.83 ms instead of 0.25 ms, and 99% of all measurements stayed below 1.5 ms.

Ceph: where the time really goes

On Ceph the picture is different. Reads are fast: on a current host with a 100 Gbit link, a single disk reads 2.7 GB/s, and 4.2 GB/s with four parallel jobs. A single write without cache, however, takes around 2 ms, compared with 0.1 to 0.25 ms locally.

We broke those 2 ms down. The network adds just 0.06 ms of round-trip time, and the SSDs in the cluster write in 0.03 ms. The time is spent in the Ceph software: every write goes to one server, which passes it on to two others and only replies once all three copies are safely stored. That is the price of redundancy. Neither debug settings nor CPU power-saving states changed anything in our tests.

This is where the cache helps. How QEMU accesses the disk makes a big difference on Ceph:

Cache mode on CephRandom write (4K, parallel)Single 4K writeSequential write
directsync4,000 IOPS2.07 ms177 MB/s
none44,900 IOPS2.10 ms253 MB/s
writeback65,700 IOPS0.07 ms420 MB/s

That is why all Ceph disks run with writeback. Databases stay safe: when your system explicitly asks for a write to be confirmed (fsync), the cache waits until Ceph has written all three copies.

Newsletter
New articles, tutorials and server news straight to your inbox. After confirming your email — only for new customers with no previous orders.
Unsubscribe anytime. No spam.

The cache had a ceiling

With sequential writes, every disk stalled at around 420 MB/s, no matter how many processes were writing. Without the cache we reached more than 1 GB/s in parallel. The cause was the size of the write cache: it held at most 24 MB that Ceph had not yet confirmed. After that, every further write had to wait.

Write cache sizeSequential write
24 MB (before)413 MB/s
64 MB1,001 MB/s
128 MB1,907–2,009 MB/s
256 MB1,913–2,049 MB/s

From 128 MB on, the disk tops out at about 2 GB/s and more cache adds nothing. So we use 128 MB. To keep a single VM from loading the cluster with 2 GB/s (times three copies), Ceph disks now also have the limit of 1,000 MB/s read and write. With the new cache a disk reaches it: we measure 1,003 MB/s in both directions. Our load data shows that this hardly affects anyone: over 30 hours, 99% of measurements stayed below 180 MB/s for writes and below 400 MB/s for reads.

New: "Database-optimized" disk cache

Since 18 November you can choose in the ZapX panel how your vServer confirms writes: in your vServer's settings, card "Disk cache".

  • Standard (recommended): Local disks run with none, Ceph disks with writeback. This is fastest for web servers, game servers, files and backups.
  • Database-optimized: The disk runs with directsync. Every write is stored durably before your system gets the confirmation. This helps databases that wait for an fsync after every transaction, such as MySQL, MariaDB or PostgreSQL.

How much it helps depends on the storage. We measured single 4K writes with fsync, the way a database commits a transaction:

StorageStandardDatabase-optimizedDifference
Local, RAID106,100/s9,800/s+60%
Local, RAID53,300/s3,600/s+9%
Ceph400–500/s500/sup to +25%

The trade-off: large and parallel writes get slower. On Ceph the difference is extreme, random 4K writes drop from 65,700 to 4,000 per second. So on Ceph we recommend the standard mode in almost every case. Your vServer picks up the new mode after a restart in the panel.

We tuned the network too

Between the hosts and the Ceph cluster, packets were dropped during load peaks because the switch buffers were set too small. Dropped packets have to be resent and cost time. We increased the buffer for regular traffic from about 3.5 to 8.4 MB. On the most affected link, drops fell by 72%, without any interruption.

Info

What you need to do now. On local NVMe storage (Dallas, Ashburn, Los Angeles): nothing, the new limit already applies there. On Ceph (Eygelshoven), the larger cache only takes effect once your VM is restarted. To apply it, restart your vServer once in the web interface: the “Restart” button, or “Stop” and then “Start”. A reboot from inside the operating system (e.g. reboot) is not enough, because the virtual machine keeps running on the host with its old settings.

Need even more? Vote on a Boost Mode

For short peaks like large imports, backups or builds, we are considering a Boost Mode: up to 2 GB/s read and write per disk on request. Nothing is planned yet. You decide whether it is worth it: vote on the suggestions board and tell us what you would use it for.

Conclusion

The biggest bottleneck was not the hardware but two settings: a per-disk limit that was much tighter than necessary, and a write cache on Ceph that was too small. We changed both, the limit even without a restart. Local NVMe storage stays faster for single synchronous writes, while Ceph keeps your data redundant. With the cache, Ceph comes close for almost every kind of access.

View vServers

Frequently asked questions

How fast is my vServer's disk?

Up to 1,000 MB/s read and write per disk. Local NVMe storage reaches that fully, on RAID5 a single write stream reaches about 640 MB/s. On Ceph, a disk reads up to the limit, and with the new cache it reaches the write limit as soon as your VM has been stopped and started once.

Do I need to restart my vServer?

Only on Ceph (Eygelshoven location). Restart it once in the web interface, or stop it and start it again. A reboot inside the operating system is not enough, because the virtual machine simply keeps running on the host. On local NVMe storage you don't need to do anything.

Is my data safe with the write cache?

Yes. When your system explicitly confirms a write (fsync), for example a database, the cache waits until all three copies are stored in Ceph. Only writes that have not been confirmed yet can be lost if the host crashes. That is why databases confirm important writes with fsync anyway.

Is "Database-optimized" worth it for my vServer?

On local NVMe storage with RAID10, yes, if it runs a database with many small transactions: up to 60% more committed writes per second. On RAID5 the gain is small, and on Ceph you are better off with the standard mode.

Why is there a limit at all?

Several VMs share the SSDs of a host or the Ceph cluster. The limit makes sure a single VM cannot slow down the others. At 1,000 MB/s in both directions it is high enough that hardly anyone reaches it in daily use.

Is RAID5 slower than RAID10?

With parallel writes and for reads, both reach our limit. A single write stream on RAID5 manages about 640 instead of up to 1,000 MB/s. Single writes that need immediate confirmation are slower: 0.25 instead of 0.10 ms inside the VM. That mainly affects databases with very many small transactions.

Eric

About the author

Eric

Developer, FiveM

Developer and system administrator at ZAP-Hosting, responsible for FiveM and RedM, as well as technical and security topics.

VPS Hosting

Full root control, NVMe SSD, and guaranteed resources.

Root accessFrom €7.29/mo.NVMe SSD
Get VPS hosting now

Feedback

Was this blog post helpful?

Your feedback helps us prioritize and improve blog articles.

Discussion

Comments

We'd love to hear your thoughts. Comments appear after a brief review.

No comments yet. Be the first to comment!

From category

More about Insights

More posts on the same topic—curated for you from our blog.

All articles in Insights
Insights7 min17 Nov 2026

FiveM: Game Server or VPS? What Our Customers Actually Use

Roughly one in five FiveM projects with us runs on a VPS. Our data shows how much RAM FiveM servers really need, how many players are online, and when each option makes sense.

Read more

Explore more

More interesting articles

Discover the latest news, tutorials, and tips from other categories.

All articles
News7 min18 Nov 2026

GTA 6 reviews are in: a 99 that ties a 28-year-old record

Only one game had ever scored 99 on Metacritic, and that was back in 1998. GTA 6 just matched it. Here is what the critics who played it early loved most.

Read more

Browse the highlights