12.15.06
Posted in programmer productivity, review at 4:58 pm by ducky
I found another paper that has some data on variability of programmers’ productivity, and again it shows that the differences between programmers isn’t as big as the “common wisdom” made me think.
In An Exploratory Study of How Developers Seek, Relate, and Collect Relevant Information during Software Maintenance Tasks, Andrew Ko et al report on the times that ten experienced programmers spent on five (relatively simple tasks). They give the average time spent as well as the standard deviation, and — like the Curtis results I mentioned before. (Note, however, that I believe Ko et al include programmers who didn’t finish in the allotted time. This will make the average lower than it would be if they waited for people to finish, and make the standard deviation appear smaller than it really is.)
| Task name |
Average time |
Standard deviation |
std dev / avg |
| Thickness |
17 |
8 |
47% |
| Yellow |
10 |
8 |
80% |
| Undo |
6 |
5 |
83% |
| Line |
22 |
12 |
55% |
| Scroll |
64 |
55 |
76% |
These numbers are the same order of magnitude as the Curtis numbers (which were 53% and 87% for the two tasks).
Ko et al don’t break out the time spent per-person, but they do break out actions per person. (They counted actions by going through screen-video-captures, ugh, tedious!) While this won’t be an exact reflection of time, it presumably is related: I would be really surprised if 12 actions took longer than 208 actions.
I’ve plotted the number of actions per task in a bar graph, with each color representing a different coder. The y-axis shows the time spent on each task, with negative numbers if they didn’t successfully complete the task. (Click on the image to see a wider version.)
Looking at the action results closely, I was struck by the impossibility of ranking the programmers.
- Coder J (white) was the only one who finished the SCROLL task (and so the only one who finished all five), so you’d think that J was really hot stuff. However, if you look at J’s performance on other tasks, it was solid but not overwhelmingly fast.
- Coder A (orange) was really fast on YELLOW and UNDO, so maybe Coder A is hot stuff. However, Coder A spun his or her wheels badly on the LINE task and never finished.
- Coder G (yellow) got badly badly stuck on LINE as well, but did a very credible job on most of the other tasks, and it looks like he or she knew enough to not waste too much time on SCROLL.
- Coder B (light purple) spent a lot of actions on THICKNESS and YELLOW, and didn’t finish UNDO or SCROLL, but B got LINE, while A didn’t.
One thing that is very clear is that there is regression to the mean, i.e. the variability on all the tasks collectively is smaller than the variability on just one task. People seem to get stuck on some things and do well on others, and it’s kind of random which ones they do well/poorly on. If you look the coders who all finished THICKNESS, YELLOW, UNDO, and SCROLL, the spread is higher on the individual tests than the aggregate:
- The standard deviation of the number of actions taken divided by the average number of actions is 36%, 69%, 60%, and 25%, respectively.
- If you look at the number of actions each developer takes to do *all* tasks, however, then the standard deviation is only 21% of the average.
- The most actions taken (“slowest”) divided by the least actions (“fastest”) is 37%, 20%, 26%, and 54% respectively.
- The overall “fastest” coder did 59% as many actions as the “slowest” overall coder.
It also seemed like there was more of a difference between the median coder and the bottom than between the median and the top. This isn’t surprising: there are limits to how quickly you can solve a problem, but no limits to how slowly you can do something.
What I take from this is that when making hiring decisions, it is not as important to get the top 1% of programmers as it is to avoid getting the bottom 50%. When managing people, it is really important to create a culture where people will give to and get help from each other to minimize the time that people are stuck.
Permalink
12.07.06
Posted in programmer productivity at 9:39 pm by ducky
I’ve heard a few people say that the most productive programmers are about 100x more productive than the least, but that didn’t seem to ring true to me. The only times that I’ve seen someone taking five months to produce as much as someone else could do in one day, it was a management problem. The low-productivity person was either goofing off or (more commonly) they were working on the wrong thing.
In Joelonsoftware, Joel reports (using data from Professor Stanley Eisenstat at Yale) that he sees about a 3x or 4x difference in the time students put into an assignment (with no effect of grades). However — and this is a big “however” — Prof. Eisenstat depends upon self-reporting from the students. Self-reporting is notoriously inacurate.
I have found it surprisingly hard to find good academic literature that measures productivity. I finally found a 1966 study by Sackman that showed a 28x difference. That number apparently got widely circulated, despite some sloppiness in its methods (as refuted by Dickey in 1981). A response to the concern by Curtis (1981) gave data that showed of about 8x and 13x min:max for two problems if and only if you tossed out one outlier who got stuck and didn’t finish that problem, timing out instead. Keeping the outlier, the 13x changed to 22x.
The outlier made me realize that the difference between the “best” programmer and the “worst” programmer is the wrong measure. You can make that number arbitrarily large by finding the right person to compare against. For example, I will be infinitely more productive than an infant. I will be infinitely more productive than someone who sleeps at his/her desk for their entire workday every day.
I would be far more interested in the standard deviation, or perhaps the productivity difference between the median and the top performer. What is the performance hit if I hire people within two standard deviations of the mean instead of trying for the ones who are better than two sigma above the mean? (This should looked at in conjunction with the cost hit of hiring the absolute best.)
This graph shows the time that it took to complete some tasks, and the number of people who took that long. (One person took between 65 and 70 minutes to complete Task 2, for example.)
Using Curtis’ data, I get that the median coder takes 2x or 3x the time that the top coder does for each of Curtis’ two tasks. (It’s not exact because the data was bucketed into five-minute intervals.) Interestingly, this is the same order of magnitude that Prof. Eisenstat found.
Note, however, that Curtis gives measures for one task. Because of regression to the mean, the variation is likely to decrease if you look at a lot of tasks. (Curtis even mentions that the outlier on Task #2 did fine on other tasks.)
The shapes of the curves are interesting: they are skewed to the left, which makes sense: there are limits to how much faster you can get, but not to how much slower you can get. They are also bimodal. This is more intriguing and mysterious, especially since Saeed and Bornat found that grades in intro CS classes are bimodal.
While this data comes from 1981, it’s the best I could find so far. (Please tell me if you have better data!) It’s distressing that we don’t have lots of good data on this — it seems like such fundamental data! However, it is really hard to do good studies like this. Ideally, you’d want a large number of programmers to all work on the same large number of complex problems. Unfortunately, it’s very difficult to get a large number of programmers to spend a large number of hours on non-useful problems.
The big message that I take from this data for my personal use is don’t get stuck. That’s easier said than done, I know. But look at it this way: if I learn new tricks for squeezing out incremental gains — like memorizing all the keyboard shortcuts — that isn’t going to help me nearly as much as not getting stuck. Keyboard shortcuts might give take me from 10 minutes to 9 minutes, but getting stuck can easily take me from 10 minutes to 100 minutes.
Now all I have to do is figure out how to not get stuck.
Permalink
