At IBM, they rank all of their scientists from top to bottom.

One of my favorite examples by Edwards Deming is a factory that has three production lines, each manned by an employee, each producing the same product on the same equipment. Furthermore, assume that there are no defects in the equipment, the manufacturing processes are under statistical control (negligible products appearing 2 standard deviations from the mean in any measured variable, iirc), and that the employees are equally skilled at their jobs. The company has a policy to promote top performers and dismiss low performers.

Every year, at the annual review, the company policy demands that they promote one average worker, keep another average worker in production, and fire the other average worker, simply because even a statistically controlled process produces some variation. This is a random process in that all three candidates have been defined as equally qualified.

In a typical software organization, where there is hardly such a thing as a controlled process, performance reviews are pure fiction if they purport to say anything about the individual. It's simply amazing hubris to think that you can attribute any given success or failure to a quality of the individual employee when your processes are not under statistical control. It is therefore amazing naivete to think that promotions in an organization of any size are handed out based on merit. They are handed out based on, at best, a record of performance, which is subject to the vagaries of chance.

Note that I'm not advocating for any particular Utopic solution. I'm just acknowledging that local reviews of system components (employees, in this case) will never produce a globally optimum solution.

Welch's rank-and-yank has been documented as, almost without debate, a failure. GE still uses some form of it, but not the draconian 10%.

The first problem with it is that it's difficult to allocate the firings across divisions. If you do this unevenly, workers and managers in the divisions hardest hit get angry about the perceived injustice. If you do it evenly, you end up firing people in high-performing divisions who might be good employees by the company-wide standard.

The second issue is that a company's proportion of deadwood decreases with successive firings. The first iteration, most of the people let go are of negative value to their teams, and few people miss them. Second iteration, some mediocre performers are fired, but managers would rather not let them go because they're still of positive value to the team, and their firing means that those remaining have to do more work. Third iteration, it's impossible to find enough deserving people, and hell breaks loose: managers get pissed off about being forced to can good people, and line workers get angry when they see respected colleagues depart.