forex vps
Trading and low-latency VPS
Distance to the matching engine, and nothing between you and the wire.
What we would deploy
€63.63/mo
€90.90−30% annual
For algorithmic execution the tail matters far more than the mean. A server that is fast on average but occasionally pauses for four milliseconds will cost you more than one that is consistently a little slower. We sell the consistency: pinned cores, single-NUMA placement, a tickless kernel and a network path chosen for stable p99 rather than headline throughput.
Why this configuration
NITRO with isolcpus and NUMA affinity removes almost every source of scheduler jitter. Deploy in the city closest to your venue: London and Frankfurt for European equities and futures, New York and Ashburn for United States markets, Tokyo and Singapore for Asian sessions.
What we would deploy
Sizing
One pinned core per feed handler, one per strategy, two left for the OS and the network stack. A single-venue crypto book with three strategies fits in six NITRO cores. Consolidated equity feeds want more: budget a core per gigabit of unfiltered market data. RAM follows book depth and retained history — 32 GB holds a full L2 book with room for a day of ticks. Do not buy cores you will not pin.
Where p99 jitter comes from
Mean latency on a modern kernel is fine. The tail is where money goes. In order of contribution: scheduler migration off your hot core, interrupts landing on it, C-state exits after an idle gap, and page faults on cold memory. Fix them in that order. isolcpus with nohz_full and rcu_nocbs on the strategy cores, irqaffinity pinned to housekeeping cores, cpuidle capped at C1, and the hot path pre-faulted and mlocked at start-up. Measure with cyclictest first, so you know the floor before your own code adds to it. NITRO cores are pinned, so there is no steal time underneath you.
Proximity, measured properly
Light in fibre runs about 5 µs per kilometre, and the installed path is typically 1.4 to 1.6 times the straight line. Frankfurt to Amsterdam is roughly 5 ms round trip. Frankfurt to Tokyo is closer to 240 ms. No amount of tuning recovers that, so the venue's metro decides the box. Frankfurt for Eurex and Deutsche Börse, London for LSE and FX, New York for the northern New Jersey venues, Tokyo and Singapore for the large crypto matching engines. CrestVPS holds its own cages in all five, which keeps the cross-connect path short and the variance low.
What goes wrong
- Tuning the kernel and leaving C-states alone. A core that idles for 200 µs between bursts exits C6 and pays several microseconds on the next packet. Cap at C1 and keep the core warm.
- Isolating cores but leaving network IRQs on them. isolcpus without matching irqaffinity and RPS settings just moves the interrupt storm onto the strategy core. Pin queues to housekeeping cores explicitly.
- Optimising microseconds on a box in the wrong metro. Fibre costs roughly 5 µs per kilometre and the real path is longer than the map. No kernel work recovers a cross-continent hop.
Tune the machine
- CPU scheduling — Pinned + isolcpus + tickless · €18
- CPU scheduling — Single-NUMA-node affinity · €14
- Everything else — Second-resolution metrics + alerting · €5
- Backups — Daily backup, 7 restore points · €4
More
Built for specific jobs
Game server hosting
Tick rate is a single-thread problem. Everything else is noise.
Private VPN and proxy endpoints
Your own exit, in a jurisdiction you picked on purpose.
Seedboxes and bulk storage
Terabytes that stay cheap and a port that stays open.
Kubernetes worker nodes
Cheap per core, dense, and identical every time.
AI inference and fine-tuning
A whole GPU, passed through, on a commitment that makes the maths work.
Hosting resellers
Your brand on the panel, our hardware underneath.