The Lie We Tell Ourselves About "Unlimited" Storage

It is 2012, and the word "Cloud" has officially lost all meaning. Every hosting provider from Oslo to Frankfurt is slapping the label on legacy infrastructure, selling you "unlimited storage" that turns out to be a partitioned slice of a 7,200 RPM SATA drive shared with three hundred other noisy neighbors. I have seen it happen too many times: a perfectly optimized Magento installation hits a wall, not because of CPU load or RAM exhaustion, but because the disk I/O wait (iowait) spikes to 40% every time a customer tries to checkout.

As systems architects, we need to stop buying marketing buzzwords and start buying IOPS (Input/Output Operations Per Second). If you are running a database-heavy application in Norway, the latency between your server and the storage backend is the single most critical metric you are likely ignoring.

The Architecture of Slowness: Centralized SAN vs. Local Storage

Most "Cloud" VPS providers utilize a centralized SAN (Storage Area Network). Your compute node is here, and your data is over there, connected via Fibre Channel or iSCSI. In theory, this offers redundancy. In practice, when a neighboring VM decides to run a massive backup job, your read/write speeds tank. The network link becomes saturated, and your MySQL queries start piling up in the process list.

At CoolVDS, we took a different architectural approach. We recognized that for raw performance, physics still applies. Distance equals latency. That is why we champion local RAID-10 SSD storage directly attached to the hypervisor.

Pro Tip: Never trust the host's promised specs. Always run your own diagnostics. If you see high "steal" time in top or consistent await times over 10ms in iostat, move your data immediately.

Diagnosing the Bottleneck

Let's get our hands dirty. You suspect disk latency is killing your app. How do you prove it? On a standard CentOS 6 or Ubuntu 12.04 LTS box, top is not enough. You need iostat (part of the sysstat package).

# Install sysstat if you haven't already
yum install sysstat -y

# Run extended statistics, updating every 2 seconds
iostat -x 2

Look at the %util and await columns. If %util is near 100% and your throughput is low (e.g., under 10MB/s), you are capped by the physical limitation of the spinning rust underneath your VM.

The "DD" Litmus Test

While not a perfect benchmark for random I/O, a simple dd command can reveal if you are on a choked storage array. Run this during peak hours:

# Test Write Speed (Bypassing Buffer Cache)
dd if=/dev/zero of=testfile bs=1G count=1 oflag=direct

If you are getting anything less than 80 MB/s, you are likely on a shared SATA backend. On our CoolVDS KVM instances with SSD backing, we consistently see write speeds dwarfing these legacy metrics, often saturating the bus bandwidth.

Optimizing MySQL 5.5 for SSD

Moving to Solid State Drives (SSD) requires retuning your database. The default my.cnf shipped with RHEL/CentOS 6 assumes rotating platters. It tries to minimize disk seeks because seeks are expensive on mechanical drives. On SSDs, random access is nearly instant.

Here is a configuration block I deployed last week for a high-traffic client to leverage the random I/O capabilities of SSDs:

[mysqld]
# Default is usually 200. On SSD, we can crank this up significantly.
innodb_io_capacity = 2000

# Disable the "neighbor" flushing meant for spinning disks.
innodb_flush_neighbors = 0

# Ensure we aren't double-buffering with the OS.
innodb_flush_method = O_DIRECT

# Vital for write-heavy workloads, though check ACID compliance requirements.
# 1 = safest, 2 = faster, 0 = fastest (risky).
innodb_flush_log_at_trx_commit = 2

Virtualization Matters: KVM vs. OpenVZ

In 2012, many budget hosts still rely on OpenVZ. OpenVZ is container-based virtualization that shares the host kernel. It is efficient, but it has a fatal flaw: lack of resource isolation. One user's heavy I/O affects everyone else.

This is why CoolVDS standardizes on KVM (Kernel-based Virtual Machine). KVM allows us to allocate dedicated block devices to your instance. Your kernel is your own. Your memory is your own. And crucially, your I/O scheduler (like CFQ, Deadline, or Noop) can be tuned specifically for your workload without the host node overriding it.

Tuning the Linux Scheduler for Virtualized Storage

Inside your VM, the default scheduler is often cfq (Completely Fair Queuing). For a virtualized guest on high-speed storage, noop or deadline is often superior because the hypervisor is already handling the sorting.

# Check current scheduler
cat /sys/block/vda/queue/scheduler
# Output: [cfq] deadline noop

# Switch to noop on the fly
echo noop > /sys/block/vda/queue/scheduler

Add this to your /etc/rc.local or kernel boot parameters to make it persistent.

Data Sovereignty: The Norwegian Advantage

We cannot discuss infrastructure without discussing jurisdiction. With the current scrutiny on the US Safe Harbor framework, housing your data within the EEA (European Economic Area) is becoming a compliance necessity, not just a preference. The Norwegian Data Protection Authority (Datatilsynet) enforces strict adherence to the Personal Data Act.

Hosting outside of Norway introduces legal complexity and network latency. If your customers are in Oslo, your data should be in Oslo. Packets traveling via the NIX (Norwegian Internet Exchange) will always beat packets routing through London or Amsterdam. We measure latency in milliseconds, but users measure it in frustration.

Summary: The TCO of Cheap Hosting

Saving 50 NOK a month on a budget VPS costs you thousands in lost productivity when you spend days debugging "slow performance" that turns out to be hardware contention. High-performance hosting is about predictability.

If you are tired of watching the iowait counter tick up while your users wait for a page load, it is time to rethink your storage strategy. Stop optimizing code for slow hardware. Fix the hardware first.

Ready to see what raw I/O throughput looks like? Deploy a CoolVDS KVM instance today and run your own benchmarks. Speed is the only metric that matters.