Skip to main content

What is Machnet?

Machnet is an open-source, DPDK-based networking stack that provides kernel-bypass messaging between cloud VMs. It enables distributed applications like databases, trading systems, and data pipelines to achieve sub-100 microsecond tail latency at high message rates — without requiring any DPDK expertise.

Build

Key Results

MetricValue
Throughput750,000 req/s (1KB messages)
P99.9 Latency61 microseconds
PlatformAzure F8s_v2 VMs

Why Machnet?

Traditional kernel networking stacks add milliseconds of overhead through context switches, buffer copies, and interrupt handling. DPDK bypasses the kernel entirely, but requires deep expertise and forces applications to link against complex libraries.

Machnet solves this by running as a sidecar process:

  • No DPDK expertise needed. Use a simple C API (machnet_send, machnet_recv) — no need to learn PMDs, mbufs, or ring buffers.
  • No recompilation. Your application communicates with Machnet over shared memory, not DPDK libraries.
  • Multi-tenant. Multiple applications on the same machine can share a single Machnet instance and NIC.
  • Cloud-native. Designed for and tested on public cloud VMs (Azure, AWS, GCP), not just bare metal.

Architecture

Machnet runs as a separate process on each machine. Applications attach to it over shared memory using a sockets-like API. Machnet processes across machines communicate via DPDK.

+-------------+   +-------------+   +-------------+
| App A | | App B | | App C |
+------+-------+ +------+-------+ +------+-------+
| | |
| Shared Memory | Shared Memory | Shared Memory
v v v
+------------------------------------------------------+
| Machnet Process |
| Channel Mgr | Flow Mgr | Packet Engine |
+------------------------------------------------------+
|
| DPDK (kernel-bypass)
v
+------------------------------------------------------+
| NIC (Cloud or SmartNIC) |
+------------------------------------------------------+

Supported Platforms

Cloud / HardwareNICDPDK Driver
Azure F8s_v2ConnectX-4 Lxnetvsc
AWS c5.xlargeENAena
GCPgVNICgvnic
Bare metalIntel E810ice / iavf
Bare metalBluefield-2mlx5
Bare metalConnectX-5/6mlx5

Next Steps

Published at ASPLOS 2026

Machnet has been accepted at ASPLOS 2026 — the premier conference on Architectural Support for Programming Languages and Operating Systems. ASPLOS is one of the most selective and prestigious venues in computer systems research.

Read the paper on ACM Digital Library