Mizar is a large scale and high-performance cloud network to run virtual machines, containers, and other compute workloads. Mizar offers flexible in-network processing, which simplifies the programming of data-plane to scale compared to traditional flow-based programming models. Unlike traditional networking solutions, Mizar relies on the natural partitioning of a cloud network to scale.
We built Mizar from ground-up on top of XDP. Mizar's main building block is an XDP program that runs on each host. The program implements virtual functions including overlay switching, routing, virtual endpoints, load-balancing, NAT, etc.
Mizar network has the following advantages:
- Support large scale of network endpoints in one cluster
- Fast network provisioning
- High network throughput and low latency
- Extendable data plane
- Unified data plane for VM and container
- Multi-tenant isolation for traffic and address spaces
We think of Mizar as a server-less platform for networking functions, in which developers extend it with capabilities without compromising performance or scale. The following diagram illustrates Mizar's high-level architecture:
Mizar's data-plane provides high performance and extensible packet processing pipeline and functions that achieve Mizar's functional, scale, and performance goals. Mizar's management-plane programs the data-plane by translating typical networking related APIs and resources to Mizar specific configuration. The programmability of the data-plane involves loading and unloading network functions at various stages of the packet processing pipeline. Mizar management-plane can support different networking technologies. The data-plane has been successfuly integrated as well in other implementations of management-planes.
Unlike traditional networking solutions, Mizar relies on the natural partitioning of a cloud network to scale. Mizar simplifies the programming of data-plane to scale by flexible in-network processing, compared to flow-based programming models. As it primarily targets use cases of cloud-networking among virtual machines and containers, Mizar reduces the control-plane overhead of several routing and switching protocols within a cloud environment (e.g., L2 learning, ARP, BGP, etc.).
The following diagram illustrates the overall logical architecture of Mizar:
- Virtual Private Cloud (VPC) domain: A flat-network of endpoints specific to a single tenant.
- Networks within a VPC: a group of Endpoints within a VPC. An operator may identify Networks as subnets of the VPC address space or any other partitioning scheme.
- Endpoint within a Network: the group of endpoints forming a network. Endpoints of a network have IP addresses from the VPC address space and need not have IP address of one subnet.
Traditionally routing between VPCs and subnets is managed by virtual switches and routers. These mandates, for example, that endpoints belong to the same subnets, and a network of endpoints must have a subnet address. Mizar does not have this restriction.
Mizar, introduces new abstract components called Bouncers and Dividers. Bouncers and Dividers are in-network and horizontally scalable hash tables that facilitate packet delivery between endpoints. The management-plane populates the Bouncers and Dividers tables according to network domain partitioning.
Bouncers' decision domain is constrained to a network. A Bouncer holds the configuration of endpoints within a network. When a packet arrives at a Bouncer, it is expected to find the destination endpoint's host and bounce the packet back to the host. Unlike a switch - where packet forwarding is performed by L2 learning - Bouncer's configuration maintains a mapping of an endpoint within a VPC to its host. The endpoint is identified by its IP address within a VPC (VNI). Bouncers rewrite the destination IP address of the outer packet to the endpoint's host.
Dividers' decision domain is constrained to VPCs. A Divider holds the configuration of all networks within a VPC; hence it divides (shards) the traffic inside the VPC across multiple bouncers. Dividers do not maintain endpoint-to-host mapping information. When a divider receives a packet, it determines which bouncer has the host information of the destination endpoint according to the network partitioning logic and rewrites the destination IP of the outer packet to the bouncer.
This overall architecture allows - among many advantages - to accelerate endpoints provisioning, as the management plane programs a finite number of hosts designated as Bouncers instead of propagating the endpoint configuration to each host.
- Thursday at 6:00PM - 8:00PM Pacific Time (biweekly, starting from April 2nd, 2020)
Resources:
- Mizar’s Sprint Planning
Channel
- 6:00PM to 8:00PM Pacific Time
- Zoom Meeting - 6:00PM to 7:00PM Pacific Time (We are including a zoom meeting for the first hour.)