2.3. keepalived Scheduling Overview
Using Keepalived provides a great deal of flexibility in distributing traffic across real servers, in part due to the variety of scheduling algorithms supported. Load balancing is superior to less flexible methods, such as Round-Robin DNS where the hierarchical nature of DNS and the caching by client machines can lead to load imbalances. Additionally, the low-level filtering employed by the LVS router has advantages over application-level request forwarding because balancing loads at the network packet level causes minimal computational overhead and allows for greater scalability.
Using assigned weights gives arbitrary priorities to individual machines. Using this form of scheduling, it is possible to create a group of real servers using a variety of hardware and software combinations and the active router can evenly load each real server.
The scheduling mechanism for Keepalived is provided by a collection of kernel patches called IP Virtual Server or IPVS modules. These modules enable layer 4 (L4) transport layer switching, which is designed to work well with multiple servers on a single IP address.
To track and route packets to the real servers efficiently, IPVS builds an IPVS table in the kernel. This table is used by the active LVS router to redirect requests from a virtual server address to and returning from real servers in the pool.
2.3.1. Keepalived Scheduling Algorithms
The structure that the IPVS table takes depends on the scheduling algorithm that the administrator chooses for any given virtual server. To allow for maximum flexibility in the types of services you can cluster and how these services are scheduled, Keepalived supports the following scheduling algorithms listed below.
- Round-Robin Scheduling
- Distributes each request sequentially around the pool of real servers. Using this algorithm, all the real servers are treated as equals without regard to capacity or load. This scheduling model resembles round-robin DNS but is more granular due to the fact that it is network-connection based and not host-based. Load Balancer round-robin scheduling also does not suffer the imbalances caused by cached DNS queries.
- Weighted Round-Robin Scheduling
- Distributes each request sequentially around the pool of real servers but gives more jobs to servers with greater capacity. Capacity is indicated by a user-assigned weight factor, which is then adjusted upward or downward by dynamic load information.Weighted round-robin scheduling is a preferred choice if there are significant differences in the capacity of real servers in the pool. However, if the request load varies dramatically, the more heavily weighted server may answer more than its share of requests.
- Least-Connection
- Distributes more requests to real servers with fewer active connections. Because it keeps track of live connections to the real servers through the IPVS table, least-connection is a type of dynamic scheduling algorithm, making it a better choice if there is a high degree of variation in the request load. It is best suited for a real server pool where each member node has roughly the same capacity. If a group of servers have different capabilities, weighted least-connection scheduling is a better choice.
- Weighted Least-Connections
- Distributes more requests to servers with fewer active connections relative to their capacities. Capacity is indicated by a user-assigned weight, which is then adjusted upward or downward by dynamic load information. The addition of weighting makes this algorithm ideal when the real server pool contains hardware of varying capacity.
- Locality-Based Least-Connection Scheduling
- Distributes more requests to servers with fewer active connections relative to their destination IPs. This algorithm is designed for use in a proxy-cache server cluster. It routes the packets for an IP address to the server for that address unless that server is above its capacity and has a server in its half load, in which case it assigns the IP address to the least loaded real server.
- Locality-Based Least-Connection Scheduling with Replication Scheduling
- Distributes more requests to servers with fewer active connections relative to their destination IPs. This algorithm is also designed for use in a proxy-cache server cluster. It differs from Locality-Based Least-Connection Scheduling by mapping the target IP address to a subset of real server nodes. Requests are then routed to the server in this subset with the lowest number of connections. If all the nodes for the destination IP are above capacity, it replicates a new server for that destination IP address by adding the real server with the least connections from the overall pool of real servers to the subset of real servers for that destination IP. The most loaded node is then dropped from the real server subset to prevent over-replication.
- Destination Hash Scheduling
- Distributes requests to the pool of real servers by looking up the destination IP in a static hash table. This algorithm is designed for use in a proxy-cache server cluster.
- Source Hash Scheduling
- Distributes requests to the pool of real servers by looking up the source IP in a static hash table. This algorithm is designed for LVS routers with multiple firewalls.
- Shortest Expected Delay
- Distributes connection requests to the server that has the shortest delay expected based on number of connections on a given server divided by its assigned weight.
- Never Queue
- A two-pronged scheduler that first finds and sends connection requests to a server that is idling, or has no connections. If there are no idling servers, the scheduler defaults to the server that has the least delay in the same manner as Shortest Expected Delay.
2.3.2. Server Weight and Scheduling
The administrator of Load Balancer can assign a weight to each node in the real server pool. This weight is an integer value which is factored into any weight-aware scheduling algorithms (such as weighted least-connections) and helps the LVS router more evenly load hardware with different capabilities.
Weights work as a ratio relative to one another. For instance, if one real server has a weight of 1 and the other server has a weight of 5, then the server with a weight of 5 gets 5 connections for every 1 connection the other server gets. The default value for a real server weight is 1.
Although adding weight to varying hardware configurations in a real server pool can help load-balance the cluster more efficiently, it can cause temporary imbalances when a real server is introduced to the real server pool and the virtual server is scheduled using weighted least-connections. For example, suppose there are three servers in the real server pool. Servers A and B are weighted at 1 and the third, server C, is weighted at 2. If server C goes down for any reason, servers A and B evenly distributes the abandoned load. However, once server C comes back online, the LVS router sees it has zero connections and floods the server with all incoming requests until it is on par with servers A and B.
