During the last decade, a growing number of services have been supported by world wide web servers. The web protocols are not only used to deliver static documents, they are also used to deliver streaming music or video. They also enable clients to use interactive applications including games or productivity applications. These services and applications have more stringent performance requirements than the delivery of static documents. Many researchers and companies have proposed solutions to improve the performance of web services and protocols during the last decade [KR2001]_ [WBK2014]_. We discuss a subset of them in this section.
A first way to improve the performance of the web protocols is to tune the servers that provide content. In the early days, documents were stored on a single server. Clients established TCP connections to this server to retrieve each document. This architecture evolved in several directions. A first way to speedup web services is to avoid unnecessary transmissions. Thanks to the `HEAD` method and the `If-Modified-Since:` header, web browsers can verify that they have the most recent version of a document in their cache.
Caches can also be used inside the network. To understand their benefits, let us consider an SME with a dozen of employees that are connected to the Internet through a low-speed link. These employees often access similar web sites. Consider that Alice and Bob want to browse today's local newspaper. Their browsers will both retrieve the newspaper's website through the low bandwidth link and store the main documents in their cache. Unfortunately, the same information passes twice over the low-speed link. Some companies have deployed web proxies to cope with this problem. A web proxy is a server that resides in the enterprise network. All the employee's browsers are configured to send their HTTP requests to this proxy. When such a proxy receives a request, it checks whether the content is already stored inside its own cache. If so, it returns it directly. Otherwise, the request is sent to the remote server and the information is stored in the proxy cache. By reducing the number of web objects that are exchanged over low-speed links, such proxies can significantly improve performance. Some companies also use them to control the websites that are contacted by their employees and sometimes block illegitimate accesses.
Proxies can also be located in front of servers. In this case, they are called reverse-proxies. Consider a dynamic web server that produces web pages by assembling information stored in different databases. When this server receives a request, it must send multiple queries to its databases and then create the HTML document. These queries and the creation of the HTML document take time and this limits the number of requests that our server can sustain. Many content providers would place a reverse proxy in front of such a server. The DNS servers are configured to point to the reverse proxy. Upon reception of a request, the reverse proxy first checks whether the response is already stored in its cache. If so, it can return it to the client without interacting with the official server. Otherwise, the reverse proxy contacts the server and then returns the response to the client.
These reverse proxies can also be used to spread the load among different servers. In the above example, consider that a server needs 10 milliseconds to process each request and that it must handle them sequentially. Such a server cannot support more than 100 requests per second. If the service becomes popular, then the content provider will need to deploy several servers. These servers could serve the same reverse proxy.
When a web user interacts with `www.service.net`, she expects that all the information comes from the `www.service.net` server. If the service is popular, there are probably tens, hundreds, thousands or more physical servers that support this service. Still, the user has the illusion that she is interacting with a single server. Several techniques have been deployed by content providers to scale web services. Consider a simple service that serves text documents from `N` different servers. There are different ways to architect such a service.
A first approach is to store all files on each physical server and rely on the DNS to distribute the load among them. Each physical server has its own IP address and when the DNS server receives a query for `www.service.net`, it returns the IP address of one of them. Some DNS servers use Round-Robin to return one of these IP addresses. Others measure the load of the physical servers and return the address of the less loaded one. Another possibility is to locate the physical servers in different regions and configure the DNS server to return the IP address of the server that is geographically closer to the client's IP address.
A second approach is to rely on `k` reverse proxies and `N-k` servers. The servers store the content and the proxies cache the most frequently used files. The proxies can be geographically close to the clients while the servers can reside in the datacenters of the content provider. The DNS server can also distribute the load among the different proxies or return the geographically closest proxy. An important point to note about reverse proxies is that they receive HTTP requests from clients and send HTTP requests to the original servers that host the content. Several companies, usually called Content Distribution Networks, have deployed such reverse proxies throughout the world to cache web content next to the end-users. A good description of such a CDN may be found in [NSS2010]_.
A second way to improve the web performance is to reduce the time required to retrieve web objects. While the first web servers returned an HTML documents with possibly a few images, today's rich web servers return one HTML document with associated style sheets, javascript code, images, fonts, ... Some of these web objects come from the original server while others are hosted on different servers. Today, a typical web page contains almost 2 MBytes of data on average. The size of the web pages continues to grow according to statistics collected by `httparchive.org`. Web pages targeted to mobile devices are slightly smaller.
A closer look at the average web page shows that it contains, on average, 27 KBytes of HTML, 120 KBytes of fonts, 60 KBytes of CSS information, almost 1 MBytes of images and more than 400 KBytes of javascript.Each of these web page requires about 70 different HTTP requests. In other words, a browser needs to send on average 70 requests to retrieve a complete web page.
Two directions have been explored to improve the delivery of these web pages. The first direction is to tune the HTTP protocol. The second approach is to change the entire network stack. We will discuss this approach after having covered the entire stack.
One of the limitations of HTTP from a performance viewpoint is that the requests that are sent by a browser must be sequential. Typically, a browser requests the HTML page. Once the page has been retrieved, the browser parses it to identify all the objects that it references and requests them one after each other. The web page can only be displayed to the user once all the required web objects have been retrieved. This implies that the browser must wait until the reception of each response before sending the next request. Another possibility is to allow the browser to send multiple requests without waiting for their corresponding responses. This approach is called `pipelining` in :rfc:`7230`.
To understand the benefits of pipelining, let us consider a simple but illustrative example. A client needs to retrieve 5 web objects that are each 100 bytes. The underlying transport connection has a 1 Gbps bandwidth but a one-way delay of 100 msec. A normal HTTP/1.x client would send the first request, wait 200 msec to receive the answer, then send another request... It would need one entire second to retrieve the five web objects. This is illustrated in the figure below.
With `pipelining`, the client sends the five requests immediately and receives the five responses after 200 msec. The figure below illustrates the benefits of `pipelining`.
However, as explained in :rfc:`7230`, there is one important limitation to `pipelining`. It can only be used to serve HTTP requests that are idempotent, i.e. none of the requests must depend on any of the previous requests in the pipeline. It turned out that it was difficult for web browsers to correctly support this requirement and very few of them have implemented `pipelining` [#fpipelining]_.
Another limitation of HTTP/1.1 is that all commands and parameters are encoded as ASCII strings. Using ASCII strings makes it easy to write simple clients or debug problems by observing packets. Unfortunately, the burden is placed on servers that need to include complex parsers that accept a wide range of partially compliant implementations. Furthermore, the flexibility of the ASCII encoding has enabled some classes of security attacks on servers [CWE444]_.
To cope with these two problems, the IETF HTTP working group developed version 2.0 of HTTP. HTTP/2.0 diverges from HTTP/1.1 in two important ways. First, HTTP/2.0 relies on binary encoding which is both more compact and easier to parse. Second, HTTP/2.0 supports multiple streams, which makes it possible to simultaneously transfer different web objects over a single transport connection. Furthermore, HTTP/2.0 also compresses the HTTP headers to reduce the amount of data transferred. This technique is described in :rfc:`7541` but is not discussed in this chapter.
