The fastest network request is a request not made. Maintaining a cache of previously downloaded data allows the client to use a local copy of the resource, thereby eliminating the request. For resources delivered over HTTP, make sure the appropriate cache headers are in place:

Whenever possible, you should specify an explicit cache lifetime for each resource, which allows the client to use a local copy, instead of re-requesting the same object all the time. Similarly, specify a validation mechanism to allow the client to check if the expired resource has been updated: if the resource has not changed, we can eliminate the data transfer.

Finally, note that you need to specify both the cache lifetime and the validation method! A common mistake is to provide only one of the two, which results in either redundant transfers of resources that have not changed (i.e., missing validation), or redundant validation checks each time the resource is used (i.e., missing or unnecessarily short cache lifetime).

Leveraging a local cache allows the client to avoid fetching duplicate content on each request. However, if and when the resource must be fetched, either because it has expired, it is new, or it cannot be cached, then it should be transferred with the minimum number of bytes. Always apply the best compression method for each asset.

The size of text-based assets, such as HTML, CSS, and JavaScript, can be reduced by 60%–80% on average when compressed with Gzip. Images, on the other hand, require more nuanced consideration:

Images account for over half of the transferred bytes of an average page, which makes them a high-value optimization target: the simple choice of an optimal image format can yield dramatically improved compression ratios; lossy compression methods can reduce transfer sizes by orders of magnitude; sizing the image to its display width will reduce both the transfer and memory footprints (see “Calculating Image Memory Requirements”) on the client. Invest into tools and automation to optimize image delivery on your site.

HTTP is a stateless protocol, which means that the server is not required to retain any information about the client between different requests. However, many applications require state for session management, personalization, analytics, and more. To enable this functionality, the HTTP State Management Mechanism (RFC 2965) extension allows any website to associate and update "cookie" metadata for its origin: the provided data is saved by the browser and is then automatically appended onto every request to the origin within the Cookie header.

The standard does not specify a maximum limit on the size of a cookie, but in practice most browsers enforce a 4 KB limit. However, the standard also allows the site to associate many cookies per origin. As a result, it is possible to associate tens to hundreds of kilobytes of arbitrary metadata, split across multiple cookies, for each origin!

Needless to say, this can have significant performance implications for your application. Associated cookie data is automatically sent by the browser on each request, which, in the worst case can add entire roundtrips of network latency by exceeding the initial TCP congestion window, regardless of whether HTTP/1.x or HTTP/2 is used:

Cookie size should be monitored judiciously: transfer the minimum amount of required data, such as a secure session token, and leverage a shared session cache on the server to look up other metadata. And even better, eliminate cookies entirely wherever possible—chances are, you do not need client-specific metadata when requesting static assets, such as images, scripts, and stylesheets.

To achieve the fastest response times within your application, all resource requests should be dispatched as soon as possible. However, another important point to consider is how these requests will be processed on the server. After all, if all of our requests are then serially queued by the server, then we are once again incurring unnecessary latency. Here’s how to get the best performance:

Without connection keepalive, a new TCP connection is required for each HTTP request, which incurs significant overhead due to the TCP handshake and slow-start. Make sure to identify and optimize your server and proxy connection timeouts to avoid closing the connection prematurely. With that in place, and to get the best performance, use HTTP/2 to allow the client and server to re-use the same connection for all requests. If HTTP/2 is not an option, use multiple TCP connections to achieve request parallelism with HTTP/1.x.

Identifying the sources of unnecessary client and server latency is both an art and science: examine the client resource waterfall (see “Analyzing the Resource Waterfall”), as well as your server logs. Common pitfalls often include the following:

HTTP/2 achieves the best performance by multiplexing requests over the same TCP connection, which enables effective request and response prioritization, flow control, and header compression. As a result, the optimal number of connections is exactly one and domain sharding is an anti-pattern.

HTTP/2 also provides a TLS connection-coalescing mechanism that allows the client to coalesce requests from different origins and dispatch them over the same connection when the following conditions are satisfied:

For example, if example.com provides a wildcard TLS certificate that is valid for all of its subdomains (i.e., *.example.com) and references an asset on static.example.com that resolves to the same server IP address as example.com, then the HTTP/2 client is allowed to reuse the same TCP connection to fetch resources from example.com and static.example.com.

An interesting side-effect of HTTP/2 connection coalescing is that it enables an HTTP/1.x friendly deployment model: some assets can be served from alternate origins, which enables higher parallelism for HTTP/1 clients, and if those same origins satisfy the above criteria then the HTTP/2 clients can coalesce requests and re-use the same connection. Alternatively, the application can be more hands-on and inspect the negotiated protocol and deliver alternate resources for each client: with sharded asset references for HTTP/1.x clients and with same-origin asset references for HTTP/2 clients.

Depending on the architecture of your application you may be able to rely on connection coalescing, you may need to serve alternate markup, or use both techniques as necessary to provide the optimal HTTP/1.x and HTTP/2 experience. Alternatively, you may consider focusing on optimizing HTTP/2 performance only; the client adoption is growing rapidly, and the extra complexity of optimizing for both protocols may be unnecessary.

Bundling multiple assets into a single response was a critical optimization for HTTP/1.x where limited parallelism and high protocol overhead typically outweighed all other concerns - see “Concatenation and Spriting”. However, with HTTP/2, multiplexing is no longer an issue, and header compression dramatically reduces the metadata overhead of each HTTP request. As a result, we need to reconsider the use of concatenation and spriting in light of its new pros and cons:

In practice, while HTTP/1.x provides the mechanisms for granular cache management of each resource, the limited parallelism forced us to bundle resources together. The latency penalty of delayed fetches outweighed the costs of decreased effectiveness of caching, more frequent and more expensive invalidations, and delayed execution.

HTTP/2 removes this unfortunate trade-off by providing support for request and response multiplexing, which means that we can now optimize our applications by delivering more granular resources: each resource can have an optimized caching policy (expiry time and revalidation token) and be individually updated without invalidating other resources in the bundle. In short, HTTP/2 enables our applications to make better use of the HTTP cache.

That said, HTTP/2 does not eliminate the utility of concatenation and spriting entirely. A few additional considerations to keep in mind:

There is no single optimal strategy for all applications: delivering a single large bundle is unlikely to yield best results, and issuing hundreds of requests for small resources may not be the optimal strategy either. The right trade-off will depend on the type of content, frequency of updates, access patterns, and other criteria. To get the best results, gather measurement data for your own application and optimize accordingly.

Server push is a powerful new feature of HTTP/2 that enables the server to send multiple responses for a single client request. That said, recall that the use of resource inlining (e.g. embedding an image into an HTML document via a data URI) is, in fact, a form of application-layer server push. As such, while this is not an entirely new capability for web developers, the use of HTTP/2 server push offers many performance benefits over inlining: pushed resources can be cached individually, reused across pages, canceled by the client, and more—see “Server Push”.

With HTTP/2 there is no longer a reason to inline resources just because they are small; we’re no longer constrained by the lack of parallelism and request overhead is very low. As a result, server push acts as a latency optimization that removes a full request-response roundtrip between the client and server - e.g. if, after sending a particular response, we know that the client will always come back and request a specific subresource, we can eliminate the roundtrip by pushing the subresource to the client.

Critical resources that block page construction and rendering (see “DOM, CSSOM, and JavaScript”) are prime candidates for the use of server push, as they are often known or can be specified upfront. Eliminating a full roundtrip from the critical path can yield savings of tens to hundreds of milliseconds, especially for users on mobile networks where latencies are often both high and highly variable.

Note that even the most naive server push strategy that opts to push assets regardless of their caching policy is, in effect, equivalent to inlining: the resource is duplicated on each page and transferred each time the parent resource is requested. However, even there, server push offers important performance benefits: the pushed response can be prioritized more effectively, it affords more control to the client, and it provides an upgrade path towards implementing much smarter strategies that leverage caching and other mechanisms that can eliminate redundant transfers. In short, if your application is using inlining, then you should consider replacing it with server push.

A naive implementation of an HTTP/2 server, or proxy, may "speak" the protocol, but without well implemented support for features such as flow control and request prioritization it can easily yield less that optimal performance—e.g., saturate user’s bandwidth by sending large low priority resources, such as images, while the browser is blocked from rendering the page until it receives higher priority resources, such as HTML, CSS, or JavaScript.

With HTTP/2 the client places a lot of trust on the server. To get the best performance, an HTTP/2 client has to be "optimistic": it annotates requests with priority information (see “Stream Prioritization”) and dispatches them to the server as soon as possible; it relies on the server to use the communicated dependencies and weights to optimize delivery of each response. A well-optimized HTTP server has always been important, but with HTTP/2 the server takes on additional and critical responsibilities that, previously, were out of scope.

Do your due diligence when testing and deploying your HTTP/2 infrastructure. Common benchmarks measuring server throughput and requests per second do not capture these new requirements and may not be representative of the actual experience as seen by your users when loading your application.