Shneiderman posed the question best: "How long will users wait for the computer to respond before they become annoyed?"  Researchers say "it depends." The delay users will tolerate depends on the perceived complexity of the task, user expertise, and feedback. Variability also plays an important role in delay tolerance. Users can tolerate moderate levels of delay variability, up to plus or minus 50 percent of the mean.
A number of studies have attempted to quantify computer response times versus user satisfaction. Robert Miller found three threshold levels of human attention: 
0.1s One tenth of a second was viewed as instantaneous.
1.0s A one-second response time was needed for users to feel they were moving freely through an information space.
10s A response time below 10 seconds was required for users to keep their attention focused on the task.
Miller proposed that the ideal response time is around two seconds.
Shneiderman agreed with Miller that a two-second limit is appropriate for simple tasks , as long as the cost is not excessive. Shneiderman found that users "pick up the pace" of computer systems, that they were more productive at shorter response rates, and that they "consistently prefer the faster pace," below 1 to 2 seconds.
Although users prefer shorter response rates, the optimum system response time (SRT) depends on task complexity. Fast SRTs cause input errors while longer response times tax short- term memory and frustrate users. Users want consistency in response times.
Because surfing the web is mainly a low-complexity activity, users prefer faster response rates. Usage studies empirically confirm this need for speed; most pages are viewed for less than a second and few for more than 10 seconds. 
I talked to Dr. Ben Shneiderman, one of the leading experts on HCI, to find out more about the relationship between speed and user satisfaction on the web.
Andy King: How does speed relate to usability and success on the web?
Ben Shneiderman: Usability plays a key role in web success storiesdesign, graphics, navigation, organization, consistency, etc. all play important roles. Speed is also vital it's hard to get users to like a slow interface, and satisfaction grows with speed. Google is a good example of an excellent service that is even more valuable and appreciated because it is fast. Speed is the strongest correlate of user satisfaction.
King: Why do we prefer shorter response times?
Shneiderman: Lively interaction keeps the engagement high. For most people, wasted time, especially while just waiting for something to happen, is annoying.
King: What happens when we exceed our attention threshold (8 to 12 seconds)?
Shneiderman: Users not only grow frustrated, but they forget their next step, and have to reconstruct their intentionsoften making mistakes that only exacerbate their frustration.
King: What do you think of the flow construct for user satisfaction on the web?
Shneiderman: Rapid movement through complex sequences of actions that move users toward a desired goal contributes to the flow experience. Users should be working just at the right level of challenge, accomplishing something they desire . There is a great thrill of finding what you want, and getting it rapidly so you can move on to the next step. 
System Response Time Guidelines
Shneiderman suggests the following guidelines for system response times: 
The speed at which your pages display can affect user perceptions of the quality, reliability, and credibility of your web site. Ramsay, Barabesi, and Preece studied the effects of slow-loading pages on user perceptions of web sites.  Using delays of two seconds to two minutes (with an interval of 19.5 seconds), they asked users to rate pages on eight criteria including "interesting content" and scannability. They found that pages with delays of 41 seconds or longer were perceived to be significantly less interesting and harder to scan. Note that the pages in this study loaded incrementally.
Jacko, Sears, and Borella studied the effects of network delay and type of document on perceived usability. They found that perceived usability of web sites was dependent on the length of delay and on the media used in web documents. When delays are short, users prefer documents that include graphics. When delays lengthen, however, users prefer text-only documents because graphics are viewed as contributing to the delay. As users become more experienced , their sensitivity to delay increases , increasing the need for "delay reduction mechanisms." 
Morris and Turner found that perceived quality of experience (Shackel's utility dimension) affects the adoption rate of IT.  How users perceive the quality of a system can affect how much they will actually use it.
They found that interface "enhancements" (graphics, animation, sound, etc.) had little effect on quality of experience "although these features may be aesthetically pleasingthey do little to remove actual barriers to the users' goal attainment ."
The speed at which your pages display affects their perceived quality and reliability. Bouch, Kuchinsky, and Bhatti investigated the effects of delay on perceived QoS in order to find an acceptable QoS level for e-commerce transactions. They tested delays from 2 to 73 seconds for both non-incremental and incrementally loaded pages.  Users rated latency quality versus delay on a scale of high, average, to low (see Table 1.1).
|Quality Rating||Range of Latency: Non-Incremental Display||Range of Latency:Incremental Display|
|High||05 seconds||039 seconds|
|Average||> 5 seconds||> 39 seconds|
|Low||> 11 seconds||> 56 seconds|
The results show a mapping between objective QoS and the users' subjective perception of QoS. Pages that displayed quickly (05 seconds) were perceived to be of high quality with high-quality products. Pages that displayed slowly (> 11 seconds) were perceived to be of low quality and untrustworthy. In fact, slower pages caused some users to feel that the security of their purchases may have been compromised, and they abandoned their transactions.
Figure 1.3 shows the actual data behind Table 1.1 for the non-incremental display. This figure plots the number of low, average, and high ratings versus latency. The range where high ratings turn to low is between 8 to 10 seconds for non-incremental downloads, closely matching what Nielsen and others have found.
Users tolerated nearly six times the delay for pages that displayed incrementally, although this tolerance decreased with usage. Test subjects rated pages as "average" with delays up to 39 seconds, and "low" with delays over 56 seconds.
The researchers also tested user requirements for speed by allowing them to click "increase quality" if they found the web page delay to be unacceptable. The average tolerance was 8.6 seconds with a standard deviation of 5.9 seconds. They attribute this large deviation in acceptable download times to contextual factors like web experience and user expectations. The longer users interact with a site, the less they tolerate delays.
Users will tolerate longer delays with tasks they perceive to be computationally complex. Users expect database access or complex calculations to take longer than displays of cached or static pages. Users form a conceptual model of system performance, which influences their tolerance for delay.
Fogg et al. found that slow-loading pages reduce ease of use, which reduces credibility (or trustworthiness and expertise). Only difficult navigation was found to hurt credibility more.