Benchmarks / Compute

Compute benchmarks

CPU, memory, disk and network performance — plus steal time, run-to-run variance and the other overselling signals hiding behind the spec sheet.

Raw performance

Identical deterministic workloads on every machine — differences are the hardware, never the benchmark.

CPU · single core

MB/s · higher = better

SHA-256 over 64 KiB buffers on one core for 3 s. Cache-resident and deterministic; uses the CPU's crypto extensions where present. source →

$ providerbench run -t cpu

CPU · all cores

MB/s · higher = better

The same SHA-256 workload running on every core simultaneously for 3 s — what the machine sustains when you actually use all the vCPUs you pay for. source →

$ providerbench run -t cpu

Memory · copy bandwidth

GB/s · higher = better

Copies between two 256 MiB buffers for 2 s — far beyond any CPU cache, so this is DRAM bandwidth, not cache tricks. source →

$ providerbench run -t memory

Memory · random access

ns · lower = better

Pointer chase through a randomly permuted 64 MiB array (fixed seed). Every load depends on the previous one, so prefetching can't hide the true latency. source →

$ providerbench run -t memory

Disk · random 4K read

IOPS · higher = better

Random 4 KiB direct-I/O reads at queue depth 4 for 2 s across the 1 GiB file — the access pattern of databases and busy filesystems. source →

$ providerbench run -t disk

Network · download

Mbps · higher = better

Sequential 50 MiB downloads from speed.cloudflare.com for up to 10 s (200 MiB max), counted at the socket. source →

$ providerbench run -t network

The overselling report

On honest hardware, identical work takes identical time. These metrics catch the difference between the cores you rent and the cores you get.

Consistency · variation

CV % · lower = better

400 identical ~20 ms hashing units back to back; this is the coefficient of variation of their runtimes. On honest hardware, identical work takes identical time. source →

High spread = noisy neighbors or burst throttling.

$ providerbench run -t steal

CPU steal time

% · lower = better

Steal ticks from /proc/stat while all cores are saturated for 5 s — the kernel's own record of CPU time the hypervisor gave to other tenants. source →

Above ~2% sustained is an oversold host. Linux only.

$ providerbench run -t steal

Tail latency · p99 / median

ratio · lower = better

How much slower the worst 1% of those 400 identical units ran compared to the median. 1.0 is perfect. source →

Burstable instances show their credit cliff here.

$ providerbench run -t steal

Disk · fsync latency (p50)

ms · lower = better

Median of 50 rounds of a 4 KiB write followed by fsync — the durability path every database commit waits on. source →

Suspiciously low values can mean the storage acknowledges before data is durable.

$ providerbench run -t disk

Full comparison

Every headline metric side by side — click a provider for its full report.

Provider / plan	€/mo	CPU 1-coreMB/s	CPU all-coreMB/s	Mem BWGB/s	Seq writeMB/s	4K readIOPS	fsync p50ms	DownloadMbps	ConsistencyCV %	Steal%
VercelSandbox · 2 vCPU / 4 GB · iad1 · 2 vCPU	—	1,325best	2,647best	6.6best	1,390best	73,757best	0.03best	3,093best	10.5best	0.30best

Raw reports: data/results/ · every chart links its benchmark source