Connection multiplexing controls the behavior of HTTP 1.0 and 1.1 request switching and server TCP connection reuse. This feature decouples the client TCP and HTTP connection from the server side TCP and HTTP connection.
Multiplexing, configured through the, provides three important benefits to improve the efficacy and performance of servers.
Reduces the number of connections to the server that must be opened and closed in a given period.
Reduces the number of concurrently open connections to the server, relative to the number of open client connections.
Allows load balancing and distribution of HTTP requests across servers using any open server side connection.
For HTTP/2 and SPDY, only the third benefit applies. These newer versions of HTTP provide the functionality outlined in 1 and 2 natively.
The exact reduction of connections to servers will depend on a few factors, such as:
Client latency and time the connection is idle.
Negotiated HTTP version.
With a one-to-one mapping of client connections to server connections, the load on a server can be extreme. With multiplexing enabled, when a client closes connection, the NSX Advanced Load Balancer intercepts the client connection close event and closes the client connection, but keeps the connection open on the server side. When the next client needs to open a connection, the NSX Advanced Load Balancer opens the client side connection but forwards the client’s request over the open, idle server side connection. The statistics of new connections per second might be significantly higher on the client side (the virtual service) than the server side (the pool). The number of requests per second will be the same, barring other features such as caching.
Most of the connections to the NSX Advanced Load Balancer from a client will be in an idle state. With HTTP 1.0 and 1.1, only a single request or response can be in flight over the connection at a time. When a server sends a response, even after it has sent the response, it cannot use the connection for other purposes. It must wait till the client has received the response, which can be for a certain time with typical Internet latency. It must wait for the client to decide to send another request and for that request to be received. Actual data transmission or receipt by the server tends to be a very small percentage of time.
With multiplexing enabled, the NSX Advanced Load Balancer receives the server response, buffers the response, acknowledges receipt to the server and transmits the response to the client. In the meantime, if another client request has been received, the NSX Advanced Load Balancer can send it over the currently idle server connection. For typical HTTP sites, particularly with typical latency, the number of open concurrent connections to the servers will be significantly less than the number of open concurrent connections from clients.
Load Balance Requests
Without multiplexing: Request distribution from traffic originating from one client connection.
With multiplexing: Request distribution from traffic originating from one client connection.
When an SE receives multiple connections from a single client or from multiple clients, requests are eligible for load balancing across any server. This ensures that a server that is slow to respond to a particular request is bypassed when the next client request is received by NSX Advanced Load Balancer. With load balancing algorithms such as round robin, they would skip a server’s connection for the next request, if that connection is currently busy with a request or response. So a server processing a complicated and slow database query might receive fewer requests than other servers processing faster requests.
Connection refers to a TCP connection, and request refers to an HTTP request and subsequent response.
HTTP 1.0 and 1.1 allow only a single request or response to go over an open TCP connection at a time. Many browsers attempt to mitigate this bottleneck by opening around six concurrent TCP connections to the destination web site.
HTTP/2 uses a different connection methodology and does not use or require the connection multiplex feature.
Impact on Other Features
Multiplexing distorts the expected view of an administrator’s world. For example, viewing connection counts from clients and to servers is very different. Static load balancing algorithms such as round robin or least connections can exhibit different behavior.
Preserve Source IP — Multiplexing explicitly decouples the client connection from the equation, when sending traffic to destination servers. When Preserve Source IP is enabled along with using the Service Engine as a default gateway of the servers, HTTP traffic cannot be multiplexed.
NTLM — NTLM authentication is generally incompatible with multiplexing. Applications that rely on NTLM, such as SharePoint or StoreFront must be configured with connection multiplexing deactivated.
Persistence — Persistence is significant for multiplexing as described below.
Multiplex Plus Persistence
Multiplexing behavior changes when server persistence is enabled:
Multiplex enabled, Persistence deactivated: Client connections and their requests are decoupled from the server side of the Service Engine. Requests are load-balanced across the servers in the pool using either new or pre-existing connections to those servers. The connections to the servers can be shared by requests from any clients.
Multiplex enabled, Persistence enabled: Client connections and their requests are sent to a single server. These requests can share connections with other clients who are persisted to the same server. Load balancing of HTTP requests is not performed.
Multiplex deactivated, Persistence enabled: The NSX Advanced Load Balancer opens a new TCP connection to the server for each connection received from the client. Connections are not shared with other clients. All requests received through all connections from the same client are sent to one server. HTTP client browsers can open many concurrent connections, and the number of client connections will be the same as the number of server connections.
Multiplex deactivated, Persistence deactivated: Connections between the client and server are one-to-one. Requests remain on the same connection they began on. Multiple connections from the same client can be distributed among the available servers.
Connection multiplexing disabled is supported with pool groups.
If the application profile has connection multiplexing deactivated, that is, if the client and server connections are tied together, the fwd_close_hdr_for_bound_connections option is applicable. It defaults to
TRUE. With fwd_close_hdr_for_bound_connections as
TRUE, if the back-end server responds with a
Connection: Close header, the SE forwards this header to the client and closes the back-end connection. On receiving
Connection: Close, the client closes the front-end connection.