Network Architecture

Learn how traffic enters, exits, and traverses our platform — from our redundant Layer 3 core and Internet edge to the Arista switching fabric in every rack.

Extreme MLXe L3 Core Three-Carrier Blend Ohio IX Peering Arista Top-of-Rack

Deterministic by Design

Cloud Propeller’s network is purpose-built, carrier-grade, redundant, and predictable. The Layer 3 core handles public IP routing, upstream transit, and peering, while the Layer 2 fabric inside each rack handles east/west traffic between hosts. Client routing and security policies live inside client-controlled virtual appliances, rather than inside the shared physical core.

Architecture Overview

Designed for Continuous Network Availability

Our network architecture consists of a redundant Extreme Networks MLXe-8 Layer 3 core handling Internet edge and routing, a three-provider blended Internet service on the upstream side with direct peering at Ohio IX, and a redundant Arista top-of-rack pair inside every rack.

Each layer has one clear job and an obvious fallback when hardware fails. The sections below cover each component in detail — how clients get connectivity, the L3 core hardware that routes it, and the switching fabric that moves tenant traffic between hosts.

Direct Connectivity via Ohio IX

AS20940

AS6181

AS16509

AS31128

AS54069

AS32899

AS30081

AS10796

AS16617

AS13335

AS7106

AS8001

AS394740

AS14978

AS63157

AS19009

AS396238

AS54113

AS33495

AS2707

AS394111

AS22062

AS2731

AS13807

AS397449

AS6939

AS394280

AS14230

AS393775

AS395848

AS55077

AS63343

AS39937

AS63293

AS62866

AS25645

AS40027

AS53471

AS15081

AS64228

AS3856

AS397095

AS63027

AS25787

AS394611

AS26554

AS394828

AS32590

AS393618

AS20009

AS46231

AS40460

Carriers, Peering & Client Connectivity

Three Providers, One Blended Edge, Over 50k IPv4 Routes at < 2ms latency

Upstream is delivered as a three-provider blended Internet service with direct peering at Ohio IX layered on top.

The three-carrier blend means there is no single-carrier dependency at the physical edge — if one provider has a fault, traffic continues over the others without client intervention. Ohio IX peering shortens paths to major cloud and content networks, cutting latency and unnecessary transit for common destinations. Clients get immediate upstream redundancy by default, with tenant-level uplinks and dedicated Spectrum options available through our meet-me-rack (MMR).

Carrier Blend

100% uptime, three-provider blended Internet service
Immediate upstream redundancy by default
No single-carrier dependency at the physical edge
Designed to keep client connectivity simple and resilient

Ohio IX Peering

Direct peering at Ohio IX
Lower latency and fewer transit dependencies
Shorter paths to major cloud and content networks
Direct peering w/ Microsoft, AWS, Apple, Valve, CacheFly, CloudFlare and more

Client Uplinks

1 Gbps (burst to 5 Gbps)
10 Gbps (burst to 50 Gbps)
Up to 100 Gbps dedicated available
Dedicated Spectrum circuits via MMR — no cross-connect fees
Additional DIA & Layer 2 connectivity to all ISPs at Cologix MMR

Client Connectivity & Security

As an infrastructure provider, Cloud Propeller offers flexible options for client connectivity while leaving internal network design decisions to the client. Organizations can consume our blended Internet service, bring in a different or additional ISP, deliver their own Layer 2 connectivity over the carrier of their choice, and manage internal Layer 3 routing and security in whatever manner best fits their environment.

Most of our clients — including small and mid-sized businesses, universities, and enterprises — run Palo Alto Networks virtual firewalls, while others use platforms such as VyOS, Sophos, pfSense, Fortinet, SonicWall, and more. Hybrid deployments are also supported for clients that prefer to, or for compliancy reasons have to, place their own physical firewalls or other dedicated network appliances alongside their cloud infrastructure.

Layer 3 Core Hardware

Carrier-Grade L3 Core

At the hardware layer, the Extreme Networks MLXe-8 is a carrier-class 8-slot routing platform delivering 3.84 Tbps of switch fabric capacity, 3.2 Tbps of forwarding capacity, 2.38 billion packets per second of routing performance, and support for up to 2.4 million IPv4 routes in hardware FIB. Its redundant management, switch fabric, power, and cooling architecture makes it well suited to high-availability routed core and Internet edge roles.

Note on Brocade / Extreme Networks

Legacy Brocade branding may still appear on the hardware shown; the MLX platform is (as of 2017) part of Extreme Networks.

Extreme Networks MLXe-8 Layer 3 core routers in production, providing Cloud Propeller’s shared routed foundation for upstream transit, peering, and Internet edge connectivity.

Arista — Top of Rack

Low-Latency, Cut-Through East/West Fabric

Arista fabric carries east/west traffic between hosts and north/south traffic back toward the routed core. This layer is intentionally simple, fast, and redundant.

Every compute rack for both Mission Critical Compute (MCC) and General Purpose Compute (GPC) is built around a redundant pair of Arista DCS-7060CX2-32S top-of-rack switches. Hosts are dual-connected into that pair, and the ToR pair uplinks back into the Layer 3 core. This keeps switching local to the rack, minimizes failure domains, and preserves predictable behavior during maintenance or hardware faults.

MCC hosts connect at 4 × 100 Gbps per host; GPC runs at 4 × 10 Gbps per host. Both tiers follow the same operating model: redundant switching, deterministic forwarding, and minimal complexity in the shared physical fabric.

Redundant Fabric

Redundant Arista ToR pair in every rack
4 × 100 Gbps Switch-to-Switch MLAGs (including rack-to-rack)
Dual-connected hosts
Dual uplinks from the rack back to the core
Maintenance without rack-wide disruption

Low-Latency Switching

Cut-through forwarding for east/west traffic
Consistent line-rate behavior under load
Deterministic intra-rack and inter-rack performance
Simple switching design with limited failure domains

Tiered Host Connectivity

MCC: 4 × 100 Gbps per host
GPC: 4 × 10 Gbps per host
Same Arista switching model across both tiers
Host-side bandwidth scales with the platform tier

Layer 2 Fabric

Low-Latency Layer 2 Fabric

Cloud Propeller’s rack fabric is built on Arista DCS-7060CX2-32S switches, a 1RU high-performance top-of-rack platform delivering 6.4 Tbps of throughput. Arista’s cut-through forwarding keeps east/west latency low by design — switching latency as low as 450 ns per hop — complemented by a shared 22 MB packet buffer pool for burst tolerance under load. In practice, this allows Cloud Propeller to maintain a simple, redundant, and deterministic rack-level switching model across both Mission Critical Compute (MCC) and General Purpose Compute (GPC) environments.

Arista DCS-7060CX2-32S top-of-rack switches in production, handling east/west traffic and uplinking each rack back into the Layer 3 core.

What is “Cut-Through” Forwarding?

Cut-through forwarding reduces switch latency by reading only the portion of an incoming Ethernet frame needed to make a forwarding decision, instead of buffering the full packet first as in traditional “store-and-forward” switching. At minimum, that means reading the destination MAC address in the first 6 bytes of the frame; in fragment-free implementations, the switch may inspect the first 64 bytes to avoid forwarding runt frames while still eliminating the delay of full packet buffering.

The result is extremely low, predictable per-hop latency under load — the operating profile high-performance data centers and latency-sensitive environments depend on. East/west traffic between Cloud Propeller hosts benefits from that same low-latency fabric design.

Need to Talk Through Connectivity?

If your workloads are sensitive to carrier diversity, peering paths, routed design, or tenant uplink requirements, we are happy to walk through the network with you.

Talk to an Engineer Explore Platform Architecture

Or view the physical facilities.