hetzner / nbg-1 — CPX32

€42.23/mo · storage local-nvme · host 4115287e5837 · 2026-07-19T08:49:25Z · submitter uptimeify

Profiles

profile	grade	bound by
postgres_oltp	B	`disk.wal_fsync.iops`
timescale_ingest	A	`disk.wal_fsync.p999_us`
patroni_member	B	`cpu.stall_p999_us`
redis_sentinel	B	`cpu.stall_p999_us`
worker_probe	B	`cpu.stall_p999_us`
playwright_node	C	`host.ram_mb`
nuxt_ssr	B	`cpu.stall_p999_us`

Tail-latency reality

fsync cumulative distribution

hover the curve — snaps to measured percentiles

Points are the measured p50 / p99 / p99.9 / max — the line between them is interpolated (we store percentiles, not a full histogram). Fixed axis: a tight curve on the left is a good host; a curve swept right is an exploding tail.

marketing claim vs measured

− typical marketing claim + what p99bench measured here

−"Up to N,000 IOPS"

+ fsync p99.9 latency at QD1 0.57 ms A

−"NVMe SSD — low latency"

+ p99.9 vs median, not the mean p50 0.11 ms → p99.9 0.24 ms A

−"Dedicated vCPUs"

+ CPU steal measured under full load 0% steal

−"High-performance cores"

+ single-thread throughput 1,694 eps A

−"Fast" (short benchmark)

+ IOPS drop over 30 min sustained load 0% degradation

−"No noisy neighbours"

+ scheduler stall p99.9 0.13 ms B

Disk

bound by disk.wal_fsync.iops

WAL fsync latency ladder (ms, QD1) Every COMMIT waits on one fdatasync. p99.9 is the graded metric — the tail that stalls a database.

avg

0.2 ms

p50

0.25 ms

p99

0.36 ms

p99.9

0.57 ms A

max

9.9 ms

8K random latency p99.9 by operation (ms)

read QD1

0.24 ms

read

1.8 ms

write

4.4 ms

mixed

2.3 ms

Throughput

seq read

13,472 MB/s

seq write

9,019 MB/s

Random IOPS

read 8K

137,607 IOPS

write 8K

112,919 IOPS

mixed 8K

86,853 IOPS

fsync

4,427 IOPS

Sustained IOPS — first vs last minute of 30 min 0% degradation over 30 min

first min

125,528 IOPS

last min

126,192 IOPS

exact values & per-metric grades

metric	value	grade
`disk.wal_fsync.p999_us`	569.34	A
`disk.wal_fsync.iops`	4427.19	B
`disk.rand_read_8k_qd1.p99_us`	150.53	A
`disk.rand_read_8k.iops`	137606.86	A
`disk.rand_write_8k.iops`	112918.97	A
`disk.seq_write.bw_mbs`	9019.21	A
`disk.seq_read.bw_mbs`	13471.59	A
`disk.steady_state.degradation_pct`	0	A

CPU

bound by cpu.stall_p999_us

Throughput (events/s)

single

1,694 eps A

multi

6,658 eps

Scheduler stall ladder (ms) How long the kernel keeps a runnable thread waiting. p99.9 is graded.

p99

0.08 ms

p99.9

0.13 ms B

max

6.9 ms

Sustained throughput — first vs last minute 0% degradation · steal 0%

first min

6,640 eps

last min

6,640 eps

exact values & per-metric grades

metric	value	grade
`cpu.single_thread_eps`	1694.35	A
`cpu.scaling_efficiency`	0.982	A
`cpu.steal_pct_under_load`	0	A
`cpu.stall_p999_us`	134	B
`cpu.steady_state.degradation_pct`	0	A
`cpu.tls_verify_s`	20679.3	B

RAM

bound by ram.bw_read_mbs

Memory bandwidth (MB/s) bw read is measured outside cache (the graded metric); seq/rnd are raw STREAM-style figures.

bw read

93,094 MB/s A

seq read

263,507 MB/s

seq write

12,089 MB/s

rnd read

19,455 MB/s

rnd write

545 MB/s

exact values & per-metric grades

metric	value	grade
`ram.bw_read_mbs`	93094.45	A

Network

bound by network.loss_pct

Round-trip p99 by target (ms) Same fixed target list from every host — distance is constant, so peering and neighbour noise surface.

hetzner-fsn1

2.9 ms

hetzner-hel1

23.8 ms

hetzner-ash

97.8 ms

Packet loss by target (%)

hetzner-fsn1

0 %

hetzner-hel1

0 %

hetzner-ash

0 %

exact values & per-metric grades

metric	value	grade
`network.loss_pct`	0	A
`network.rtt_jitter_ratio`	1.0674157303370788	A