How it works
Traditional Linux OS adds 20-200 microseconds just moving packets through the kernel. Kernel bypass methods eliminate that by giving the application direct access to the NIC's receive and transmit queues. This comes at the computational cost of continuous polling. BareMetal does away with the OS tax and delivers the maximum computational resources to the application. It makes for deterministic performance with minimal variance in latency. The low-jitter, fast cold-boot performance makes BareMetal Kernel ideal for HFTs who use event-driven strategies.
Overview and pricing
We work directly with HFTs and with technology manufacturers by licensing BareMetal Kernel and requisite drivers. One-time charges may apply in case of custom driver development.
BareMetal Kernel vs Linux
Scheduling tail latency (P99.9):
• Linux: 16.760ms (+1.135ms over target)
• BareMetal: 15.631ms (+0.006ms)
Network polling:
• Linux: ~109ns
• BareMetal: 25ns — 4.4× faster
"For a decentralized system running multiple auction nodes, OS jitter means each node closes its 1-second window at a slightly different real-world moment. That inconsistency across nodes is not as egregious as traditional co-location arbitrage, but it allows for a smaller-scale structural exploit. BareMetal Kernel removes these inconsistencies and levels the playing field for the traders."
Varun Madhok
CEO, Return Infinity